not yet published
Conference article
Open Access
Classifying the ripeness of mangoes using image processing and deep learning
Jirapattarasakul S., Thanyacharoen S., Leone G. R., Akkaralaertsest T., Yingthawornsuk T.
Mango is a popular fruit that comes in many different varieties. Each variety has a different flavor, aroma, and texture. Selecting mangoes at the appropriate level of ripeness is therefore important to both consumers and producers. Problems frequently encountered include inconsistencies in sorting mangoes using traditional methods, which often rely on human experience and eyesight. This can affect the quality of mango product. This research focuses on developing a ripeness-rawness screening system for Okrong and Mahachanok mango varieties using Deep Learning and Image Processing techniques. Two CNN (Convolutional Neural Network) models, VGG16, MobileNetV2, and CNN1D, were used to analyze mango images and distinguish between ripe and raw levels. The test results showed that the VGG16 model achieved the highest performance in screening ripeness-rawness of both mango varieties, with an accuracy of 98%, followed by MobileNetV2 at 96% and CNN1D at 92%. For the ripeness-rawness classification of only the Okrong mango variety, the VGG16 model still achieved the highest performance, with an accuracy of 99%, followed by MobileNetV2 at 96% and CNN1D at 95%. These results indicate that CNN models, particularly VGG16, have great potential for application in developing automated mango sorting system based on ripeness-rawness levels. This proposed work can significantly improve efficiency in managing and selecting mango product.
Jirapattarasakul S., Thanyacharoen S., Leone G. R., Akkaralaertsest T., Yingthawornsuk T.
Mango is a popular fruit that comes in many different varieties. Each variety has a different flavor, aroma, and texture. Selecting mangoes at the appropriate level of ripeness is therefore important to both consumers and producers. Problems frequently encountered include inconsistencies in sorting mangoes using traditional methods, which often rely on human experience and eyesight. This can affect the quality of mango product. This research focuses on developing a ripeness-rawness screening system for Okrong and Mahachanok mango varieties using Deep Learning and Image Processing techniques. Two CNN (Convolutional Neural Network) models, VGG16, MobileNetV2, and CNN1D, were used to analyze mango images and distinguish between ripe and raw levels. The test results showed that the VGG16 model achieved the highest performance in screening ripeness-rawness of both mango varieties, with an accuracy of 98%, followed by MobileNetV2 at 96% and CNN1D at 92%. For the ripeness-rawness classification of only the Okrong mango variety, the VGG16 model still achieved the highest performance, with an accuracy of 99%, followed by MobileNetV2 at 96% and CNN1D at 95%. These results indicate that CNN models, particularly VGG16, have great potential for application in developing automated mango sorting system based on ripeness-rawness levels. This proposed work can significantly improve efficiency in managing and selecting mango product.
See at:
gcmm2024.rmutk.ac.th 
| CNR IRIS | CNR IRIS 
| CNR IRIS
2026
Other
Open Access
Addressing SAR ship class imbalance via targeted oversampling of foundation model embeddings
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
Foundation models pretrained on remote sensing data have shown promise for downstream tasks, yet their behaviour under class imbalance remains underexplored. We benchmark two foundation models, DOFA and SAR-JEPA, against ImageNet-pretrained models on the severely imbalanced OpenSARShip dataset. We apply four feature-space oversampling techniques exclusively to minority classes, scaling them to three times their original size. Our approach achieves up to 8.34% Macro-F1 and 7.34% accuracy improvements over baseline foundation models, demonstrating that targeted oversampling enables better balanced performance on SAR ship classification. We provide code to pre-extract embeddings, and reproducible experiments optimised for free-tier Google Colab https://github.com/cm-awais/SARShipfoundationModels.DOI: 10.5281/zenodo.22091713
Metrics:
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
Foundation models pretrained on remote sensing data have shown promise for downstream tasks, yet their behaviour under class imbalance remains underexplored. We benchmark two foundation models, DOFA and SAR-JEPA, against ImageNet-pretrained models on the severely imbalanced OpenSARShip dataset. We apply four feature-space oversampling techniques exclusively to minority classes, scaling them to three times their original size. Our approach achieves up to 8.34% Macro-F1 and 7.34% accuracy improvements over baseline foundation models, demonstrating that targeted oversampling enables better balanced performance on SAR ship classification. We provide code to pre-extract embeddings, and reproducible experiments optimised for free-tier Google Colab https://github.com/cm-awais/SARShipfoundationModels.DOI: 10.5281/zenodo.22091713
Metrics:
See at:
CNR IRIS | openreview.net | ZENODO | CNR IRIS
2026
Journal article
Open Access
A Survey on SAR ship classification using deep learning
Ch Muhammad Awais, Marco Reggiannini, Davide Moroni, Emanuele Salerno
Deep learning (DL) has become a central approach for ship classification using synthetic aperture radar (SAR) imagery. This survey reviews 74 representative studies selected from 187 publications, categorizing them into a taxonomy with four main dimensions: (i) DL architectures, (ii) datasets, (iii) image augmentation, and (iv) learning techniques. We analyze how approaches such as handcrafted feature integration, data augmentation, fine-tuning, and transfer learning influence classification performance, and summarize the use of public benchmarks including OpenSARShip and FUSARShip. This survey highlights key challenges: limited data availability, class imbalance, lack of standardized metrics, and limited interpretability of DL models. Future research directions include the development of SAR-specific DL architectures, advanced augmentation and generative approaches, integration of handcrafted and deep features, interpretable DL, and stronger interdisciplinary collaboration. By addressing these challenges, DL-based SAR ship classification can achieve greater robustness, accuracy, and transparency, ultimately strengthening maritime surveillance and operational monitoring.Source: IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
DOI: 10.1109/jstars.2026.3695704
Metrics:
Ch Muhammad Awais, Marco Reggiannini, Davide Moroni, Emanuele Salerno
Deep learning (DL) has become a central approach for ship classification using synthetic aperture radar (SAR) imagery. This survey reviews 74 representative studies selected from 187 publications, categorizing them into a taxonomy with four main dimensions: (i) DL architectures, (ii) datasets, (iii) image augmentation, and (iv) learning techniques. We analyze how approaches such as handcrafted feature integration, data augmentation, fine-tuning, and transfer learning influence classification performance, and summarize the use of public benchmarks including OpenSARShip and FUSARShip. This survey highlights key challenges: limited data availability, class imbalance, lack of standardized metrics, and limited interpretability of DL models. Future research directions include the development of SAR-specific DL architectures, advanced augmentation and generative approaches, integration of handcrafted and deep features, interpretable DL, and stronger interdisciplinary collaboration. By addressing these challenges, DL-based SAR ship classification can achieve greater robustness, accuracy, and transparency, ultimately strengthening maritime surveillance and operational monitoring.Source: IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
DOI: 10.1109/jstars.2026.3695704
Metrics:
See at:
CNR IRIS | ieeexplore.ieee.org | CNR IRIS
2026
Book
Open Access
Design manual for accessibility and the inclusive use of cultural heritage: from human functioning to the functioning of cultural places
Bernacchia Monica, Biocca Luigi, Buzzi Marina, Capirci Olga, Caponera Federica, Carella Giuseppina, Ceccarani Patrizia, Cetorelli Gabriella, Contardi Anna, Della Fina Valentina, Di Renzo Alessio, Filetici Maria Grazia, Galesi Giulio, Genta Chiara, Gervasio Carmine Fernando, Grassini Aldo, Guido Manuel Roberto, Leporini Barbara, Luongo Mariajosè, Maglorio Massimo, Marino Carmen, Mezzelani Alessandra, Montuschi Carla, Pagano Alfonsina, Papi Luca, Pennacchi Barbara, Pietroni Eva, Ricci Enrico, Rossini Mauro, Russo Francesco Paolo, Schiavone Elisabetta, Schivo Flavia, Scianna Andrea, Trasatti Annalisa, Zanut Stefano
The English-language version of the Manual, presented in an expanded edition in March 2026, represents an update of the volume with two appendices that are part of the institutional series of Policy Briefs of the Department of Human and Social Sciences, Cultural Heritage (DSU) of the National Research Council (CNR). This new edition, in addition to gathering the salient aspects of "universal design," starting from the concept of the functioning of individuals to the functioning of cultural places, presents additional current aspects through documents that summarize the results of research on relevant topics and their implications for policy and social debate. In the first addendum, the choice was to explore the issues of appropriate, respectful, and correct language regarding accessibility based on Legislative Decree 62/2024, centered on a "Person First" vision, as enshrined by the World Health Organization (2001) and the Convention on the Rights of Persons with Disabilities (2006). The second document proposes an integrated approach to accessible trails, understood as landscape experiences to be enjoyed independently and safely, capable of meeting the needs of each individual
Bernacchia Monica, Biocca Luigi, Buzzi Marina, Capirci Olga, Caponera Federica, Carella Giuseppina, Ceccarani Patrizia, Cetorelli Gabriella, Contardi Anna, Della Fina Valentina, Di Renzo Alessio, Filetici Maria Grazia, Galesi Giulio, Genta Chiara, Gervasio Carmine Fernando, Grassini Aldo, Guido Manuel Roberto, Leporini Barbara, Luongo Mariajosè, Maglorio Massimo, Marino Carmen, Mezzelani Alessandra, Montuschi Carla, Pagano Alfonsina, Papi Luca, Pennacchi Barbara, Pietroni Eva, Ricci Enrico, Rossini Mauro, Russo Francesco Paolo, Schiavone Elisabetta, Schivo Flavia, Scianna Andrea, Trasatti Annalisa, Zanut Stefano
The English-language version of the Manual, presented in an expanded edition in March 2026, represents an update of the volume with two appendices that are part of the institutional series of Policy Briefs of the Department of Human and Social Sciences, Cultural Heritage (DSU) of the National Research Council (CNR). This new edition, in addition to gathering the salient aspects of "universal design," starting from the concept of the functioning of individuals to the functioning of cultural places, presents additional current aspects through documents that summarize the results of research on relevant topics and their implications for policy and social debate. In the first addendum, the choice was to explore the issues of appropriate, respectful, and correct language regarding accessibility based on Legislative Decree 62/2024, centered on a "Person First" vision, as enshrined by the World Health Organization (2001) and the Convention on the Rights of Persons with Disabilities (2006). The second document proposes an integrated approach to accessible trails, understood as landscape experiences to be enjoyed independently and safely, capable of meeting the needs of each individual
See at:
CNR IRIS | www.dsu.cnr.it | CNR IRIS
2026
Other
Restricted
Relazione CNR-ISTI / SIMeM CNR-ISTI / SIMeM. Accordo di collaborazione scientifica - Relazione conclusiva sulle attività svolte e sui risultati di disseminazione
Martinelli Massimo
La presente relazione descrive le attività svolte nell'ambito della collaborazione scientifica tra il Consiglio Nazionale delle Ricerche - Istituto di Scienza e Tecnologie dell'Informazione "A. Faedo" (CNR-ISTI) e la Società Italiana di Medicina di Montagna (SIMeM), con riferimento al periodo aprile 2020 - aprile 2026. Sono inoltre richiamate, come contesto scientifico e tecnologico, alcune attività propedeutiche avviate nel periodo 2016-2019 e risultati collegati alla sinergia tra i rispettivi gruppi di ricerca. La collaborazione ha integrato competenze di medicina di montagna, telemedicina, sistemi informativi, gestione di dati clinico-epidemiologici, supporto alla decisione e metodi di intelligenza artificiale.
Martinelli Massimo
La presente relazione descrive le attività svolte nell'ambito della collaborazione scientifica tra il Consiglio Nazionale delle Ricerche - Istituto di Scienza e Tecnologie dell'Informazione "A. Faedo" (CNR-ISTI) e la Società Italiana di Medicina di Montagna (SIMeM), con riferimento al periodo aprile 2020 - aprile 2026. Sono inoltre richiamate, come contesto scientifico e tecnologico, alcune attività propedeutiche avviate nel periodo 2016-2019 e risultati collegati alla sinergia tra i rispettivi gruppi di ricerca. La collaborazione ha integrato competenze di medicina di montagna, telemedicina, sistemi informativi, gestione di dati clinico-epidemiologici, supporto alla decisione e metodi di intelligenza artificiale.
2026
Journal article
Restricted
Blood eosinophils are accurate biomarkers for the management of eosinophilic oesophagitis: prospective, multi‐centre study
Visaggi Pierfrancesco, Pellegatta Gaia, Siboni Stefano, Maniero Daria, Del Corso Giulio, Solinas Irene, Mitilini Mauro, Testi Federico, Cairoli Gaia, Dulmin Isabella, Poletti Valeria, Marcozzi Giacomo, Sozzi Marco, Piazza Lucia, Stefani Donati Delio, Bellini Massimo, Repici Alessandro, Savarino Edoardo Vincenzo, Bortoli Nicola De
Background & Aims: The diagnosis and follow-up of eosinophilic oesophagitis (EoE) currently rely on repeated upper endoscopies (EGD) with biopsies, which are invasive, resource-intensive and environmentally costly. Non-invasive biomarkers for EoE are needed. We investigated the role of blood eosinophils and lymphocytes in the management of EoE. Methods: This was a prospective study conducted at four EoE referral centres. Consecutive adults undergoing EGD with biopsies for known or suspected EoE were enrolled. Based on oesophageal peak eosinophil count (PEC) and clinical history, patients were divided into EoE (histologically active or in remission) and non-EoE dysphagia (NED). Prior to the EGD, a full blood count was obtained. Clinical, endoscopic and histologic findings were recorded. Receiver operating characteristic curve analysis was used to assess the predictive ability of blood biomarkers (AUC). Results: We enrolled 209 patients (123 EoE and 86 NED). For the diagnosis of EoE, an absolute eosinophil count (AEC) of 155 eosinophils/mm3 had an AUC of 85%. For the assessment of histological disease activity, an AEC of 325 eosinophils/mm3 had an AUC of 70.5% for the identification of histological remission following treatment. AEC showed a positive correlation with PEC on histology and the EoE endoscopic reference score with Spearman's Rho of 0.4 (p < 0.0001). Conclusion: Eosinophil absolute and relative counts in the peripheral blood could be used in the initial assessment of patients presenting with dysphagia to accurately differentiate EoE from NED and to predict histological remission of EoE.Source: ALIMENTARY PHARMACOLOGY & THERAPEUTICS, vol. 63 (issue 9), pp. 1256-1264
DOI: 10.1111/apt.70524
Metrics:
Visaggi Pierfrancesco, Pellegatta Gaia, Siboni Stefano, Maniero Daria, Del Corso Giulio, Solinas Irene, Mitilini Mauro, Testi Federico, Cairoli Gaia, Dulmin Isabella, Poletti Valeria, Marcozzi Giacomo, Sozzi Marco, Piazza Lucia, Stefani Donati Delio, Bellini Massimo, Repici Alessandro, Savarino Edoardo Vincenzo, Bortoli Nicola De
Background & Aims: The diagnosis and follow-up of eosinophilic oesophagitis (EoE) currently rely on repeated upper endoscopies (EGD) with biopsies, which are invasive, resource-intensive and environmentally costly. Non-invasive biomarkers for EoE are needed. We investigated the role of blood eosinophils and lymphocytes in the management of EoE. Methods: This was a prospective study conducted at four EoE referral centres. Consecutive adults undergoing EGD with biopsies for known or suspected EoE were enrolled. Based on oesophageal peak eosinophil count (PEC) and clinical history, patients were divided into EoE (histologically active or in remission) and non-EoE dysphagia (NED). Prior to the EGD, a full blood count was obtained. Clinical, endoscopic and histologic findings were recorded. Receiver operating characteristic curve analysis was used to assess the predictive ability of blood biomarkers (AUC). Results: We enrolled 209 patients (123 EoE and 86 NED). For the diagnosis of EoE, an absolute eosinophil count (AEC) of 155 eosinophils/mm3 had an AUC of 85%. For the assessment of histological disease activity, an AEC of 325 eosinophils/mm3 had an AUC of 70.5% for the identification of histological remission following treatment. AEC showed a positive correlation with PEC on histology and the EoE endoscopic reference score with Spearman's Rho of 0.4 (p < 0.0001). Conclusion: Eosinophil absolute and relative counts in the peripheral blood could be used in the initial assessment of patients presenting with dysphagia to accurately differentiate EoE from NED and to predict histological remission of EoE.Source: ALIMENTARY PHARMACOLOGY & THERAPEUTICS, vol. 63 (issue 9), pp. 1256-1264
DOI: 10.1111/apt.70524
Metrics:
See at:
Alimentary Pharmacology & Therapeutics | CNR IRIS | CNR IRIS | CNR IRIS | onlinelibrary.wiley.com | Alimentary Pharmacology & Therapeutics
2026
Other
Restricted
Spoke 9 - AGRITECH 36-MONTH REPORT
Pucci Laura, Tomassi Elena, Arouna Nafiou, Gabriele Morena, Peres Fabbri Laryssa, Pozzo Luisa, Conte Giuseppe, Cremonesi Paola, Castiglioni Bianca, Moroni Davide, Martinelli Massimo
This document represents the 36-month report on products of animal origin intended for human consumption.Project(s): Spoke 9 AGRITECH
Pucci Laura, Tomassi Elena, Arouna Nafiou, Gabriele Morena, Peres Fabbri Laryssa, Pozzo Luisa, Conte Giuseppe, Cremonesi Paola, Castiglioni Bianca, Moroni Davide, Martinelli Massimo
This document represents the 36-month report on products of animal origin intended for human consumption.Project(s): Spoke 9 AGRITECH
2026
Contribution to journal
Restricted
Blood pressure measurement in Italian mountain shelters: a world hypertension day initiative
Bilo Grzegorz, Zanotti Lucia, Croce Alessandro, D’angelo Carla, Faini Andrea, Martinelli Massimo, Muiesan Maria Lorenza, Pengo Martino F., Pratali Lorenza, Vega Deceno Jose Ivan, Soranna Davide, Zambon Antonella, Virdis Agostino, Strapazzon Giacomo, Agazzi Giancelso, Parati Gianfranco
World Hypertension Day and May Measurement Month are large-scale population initiatives aimed at increasing awareness of hypertension and providing opportunistic blood pressure (BP) screening. In Italy, “Blood Pressure Measurement in Mountain Shelters” has been conducted since 2016 as a joint initiative of Italian Society of Hypertension, Italian Alpine Club and Italian Society of Mountain Medicine, with the aim of providing advice and education on hypertension and cardiovascular risk among visitors of mountain shelters. Methods. Trained volunteers provided free advice to visitors of mountain shelters throughout Italy. Basic demographic, lifestyle and clinical information was collected with anonymous questionnaire. Seated BP was obtained using manual or oscillometric devices as the average of three measurements. The initiative took place in summer when mountain attendance is highest. The present analysis of merged data collected in 2019, 2022, 2023, 2024 and 2025 editions focuses on describing participant characteristics, including BP, stratified by altitude of data collection. Results. After exclusion of incomplete records (missing age, sex or BP), data of 12,317 participants from 140 mountain shelters were analysed. No relevant differences were observed across the different years in terms of participant characteristics or BP levels. Participant characteristics are reported in the Table. Median BP and body mass index were within normal limits, and the prevalence of diabetes, hypercholesterolaemia and hypertension was lower than that in the general Italian population. Fewer than 15% of participants reported taking at least one antihypertensive medication. Higher altitudes of data collection were associated with higher male sex prevalence and heart rate and with lower age, BMI, SpO2, prevalence of hypercholesterolemia and hypertension. Only minor differences were observed in measured BP. Conclusions. Blood Pressure Measurement in Mountain Shelters is a unique initiative, willingly attended by visitors to mountain areas. Participants generally displayed a favourable cardiovascular risk profile; however, a substantial proportion, including those assessed at higher altitudes, presented at least one cardiovascular risk factor. This screening and educational initiative represents a valuable opportunity to identify individuals at risk and to provide counselling on cardiovascular prevention among visitors of mountain shelters.Source: JOURNAL OF HYPERTENSION, vol. 44(Suppl 1), pp. 70-71
DOI: 10.1097/01.hjh.0001195860.36068.a1
Metrics:
Bilo Grzegorz, Zanotti Lucia, Croce Alessandro, D’angelo Carla, Faini Andrea, Martinelli Massimo, Muiesan Maria Lorenza, Pengo Martino F., Pratali Lorenza, Vega Deceno Jose Ivan, Soranna Davide, Zambon Antonella, Virdis Agostino, Strapazzon Giacomo, Agazzi Giancelso, Parati Gianfranco
World Hypertension Day and May Measurement Month are large-scale population initiatives aimed at increasing awareness of hypertension and providing opportunistic blood pressure (BP) screening. In Italy, “Blood Pressure Measurement in Mountain Shelters” has been conducted since 2016 as a joint initiative of Italian Society of Hypertension, Italian Alpine Club and Italian Society of Mountain Medicine, with the aim of providing advice and education on hypertension and cardiovascular risk among visitors of mountain shelters. Methods. Trained volunteers provided free advice to visitors of mountain shelters throughout Italy. Basic demographic, lifestyle and clinical information was collected with anonymous questionnaire. Seated BP was obtained using manual or oscillometric devices as the average of three measurements. The initiative took place in summer when mountain attendance is highest. The present analysis of merged data collected in 2019, 2022, 2023, 2024 and 2025 editions focuses on describing participant characteristics, including BP, stratified by altitude of data collection. Results. After exclusion of incomplete records (missing age, sex or BP), data of 12,317 participants from 140 mountain shelters were analysed. No relevant differences were observed across the different years in terms of participant characteristics or BP levels. Participant characteristics are reported in the Table. Median BP and body mass index were within normal limits, and the prevalence of diabetes, hypercholesterolaemia and hypertension was lower than that in the general Italian population. Fewer than 15% of participants reported taking at least one antihypertensive medication. Higher altitudes of data collection were associated with higher male sex prevalence and heart rate and with lower age, BMI, SpO2, prevalence of hypercholesterolemia and hypertension. Only minor differences were observed in measured BP. Conclusions. Blood Pressure Measurement in Mountain Shelters is a unique initiative, willingly attended by visitors to mountain areas. Participants generally displayed a favourable cardiovascular risk profile; however, a substantial proportion, including those assessed at higher altitudes, presented at least one cardiovascular risk factor. This screening and educational initiative represents a valuable opportunity to identify individuals at risk and to provide counselling on cardiovascular prevention among visitors of mountain shelters.Source: JOURNAL OF HYPERTENSION, vol. 44(Suppl 1), pp. 70-71
DOI: 10.1097/01.hjh.0001195860.36068.a1
Metrics:
See at:
CNR IRIS | CNR IRIS | CNR IRIS | journals.lww.com
2026
Journal article
Open Access
Research progress on Ocean observations technology and information systems
Tsabaris Christos, Pieri Gabriele
The oceans play a crucial role in the global ecosystem; they shape trends in the climate, weather, water management, and health (including biogeochemical cycles). Although innovative technologies have been developed for the marine environment to better understand the ocean’s processes, several constraints remain in ocean observation, hindering the replacement of laboratory and routine monitoring methods. TSource: JOURNAL OF MARINE SCIENCE AND ENGINEERING, vol. 14 (issue 9)
DOI: 10.3390/jmse14090774
Project(s): New Approach to Underwater Technologies for Innovative, Low-cost Ocean obServation
Metrics:
Tsabaris Christos, Pieri Gabriele
The oceans play a crucial role in the global ecosystem; they shape trends in the climate, weather, water management, and health (including biogeochemical cycles). Although innovative technologies have been developed for the marine environment to better understand the ocean’s processes, several constraints remain in ocean observation, hindering the replacement of laboratory and routine monitoring methods. TSource: JOURNAL OF MARINE SCIENCE AND ENGINEERING, vol. 14 (issue 9)
DOI: 10.3390/jmse14090774
Project(s): New Approach to Underwater Technologies for Innovative, Low-cost Ocean obServation
Metrics:
See at:
Journal of Marine Science and Engineering | IRIS Cnr | IRIS Cnr | CNR IRIS
2026
Other
Open Access
Fusion of foundation models for imbalanced SAR ship classification
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
While foundation models pretrained on remote sensing data worked successfully in various advanced tasks, their susceptibility to class imbalance and the potential of targeted mitigation strategies remain underexplored. We present an empirical evaluation of six remote sensing foundation models on two SAR ship classification datasets: OpenSARShip and FUSARShip. We demonstrate that commonly used metrics such as accuracy and weighted F1-score mask poor performance on minority classes, and advocate for macro F1-score as a more reliable measure for imbalanced datasets. Our experiments reveal that several foundation models underperform ImageNet-pretrained baselines (ResNet, VGG, ViT). We investigate four fusion strategies to combine embeddings from multiple foundation models in both full and lightweight configurations, where lightweight variants exclude the two poorest-performing models. Lightweight late fusion with informed weighting achieves up to 5.5 percentage point improvements in macro F1 over the best individual foundation models, while reducing computational overhead.DOI: 10.5281/zenodo.22091558
Metrics:
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
While foundation models pretrained on remote sensing data worked successfully in various advanced tasks, their susceptibility to class imbalance and the potential of targeted mitigation strategies remain underexplored. We present an empirical evaluation of six remote sensing foundation models on two SAR ship classification datasets: OpenSARShip and FUSARShip. We demonstrate that commonly used metrics such as accuracy and weighted F1-score mask poor performance on minority classes, and advocate for macro F1-score as a more reliable measure for imbalanced datasets. Our experiments reveal that several foundation models underperform ImageNet-pretrained baselines (ResNet, VGG, ViT). We investigate four fusion strategies to combine embeddings from multiple foundation models in both full and lightweight configurations, where lightweight variants exclude the two poorest-performing models. Lightweight late fusion with informed weighting achieves up to 5.5 percentage point improvements in macro F1 over the best individual foundation models, while reducing computational overhead.DOI: 10.5281/zenodo.22091558
Metrics:
See at:
CNR IRIS | openreview.net | ZENODO | CNR IRIS
2026
Conference article
Open Access
Human vs LLM-generated alt text for STEM image accessibility: a comparative study
Cardia Marco, Angileri Letizia, Poggianti Camilla, Leporini Barbara
Image descriptions are a fundamental aspect of web accessibility, allowing blind and low-vision users to access visual information through assistive technologies such as screen readers. Despite their importance, high-quality alternative text is often missing or inadequate, especially for complex images like diagrams, graphs, and other STEM-related content. Recent advances in generative artificial intelligence and large language models (LLMs) have renewed interest in automatically generating image descriptions, raising the question of whether these systems can reliably support accessibility at scale.In this paper, we present a comparative study of human-written and LLM-generated alternative text for STEM images, focusing on accessibility-critical aspects rather than surface-level textual similarity. Using a curated dataset of images with expert-authored reference descriptions, we evaluate the outputs of state-of-the-art multimodal LLMs through a mixed-methods approach. Our evaluation combines traditional automated metrics, such as BLEU and METEOR, with a human-centered analysis targeting accuracy, informational completeness, and the presence of hallucinations. The results show that while LLMs often produce fluent and seemingly informative descriptions, substantial gaps remain compared to human-written alt text, particularly in conveying formal semantics and essential structural details required for accessibility. We further observe that commonly used automated metrics only partially capture accessibility-relevant errors, underscoring the need for evaluation methodologies based on the needs of screen reader users. We discuss the implications of these findings for the design, evaluation, and deployment of generative AI systems in accessible web and educational contexts.DOI: 10.1145/3800424.3800457
Metrics:
Cardia Marco, Angileri Letizia, Poggianti Camilla, Leporini Barbara
Image descriptions are a fundamental aspect of web accessibility, allowing blind and low-vision users to access visual information through assistive technologies such as screen readers. Despite their importance, high-quality alternative text is often missing or inadequate, especially for complex images like diagrams, graphs, and other STEM-related content. Recent advances in generative artificial intelligence and large language models (LLMs) have renewed interest in automatically generating image descriptions, raising the question of whether these systems can reliably support accessibility at scale.In this paper, we present a comparative study of human-written and LLM-generated alternative text for STEM images, focusing on accessibility-critical aspects rather than surface-level textual similarity. Using a curated dataset of images with expert-authored reference descriptions, we evaluate the outputs of state-of-the-art multimodal LLMs through a mixed-methods approach. Our evaluation combines traditional automated metrics, such as BLEU and METEOR, with a human-centered analysis targeting accuracy, informational completeness, and the presence of hallucinations. The results show that while LLMs often produce fluent and seemingly informative descriptions, substantial gaps remain compared to human-written alt text, particularly in conveying formal semantics and essential structural details required for accessibility. We further observe that commonly used automated metrics only partially capture accessibility-relevant errors, underscoring the need for evaluation methodologies based on the needs of screen reader users. We discuss the implications of these findings for the design, evaluation, and deployment of generative AI systems in accessible web and educational contexts.DOI: 10.1145/3800424.3800457
Metrics:
See at:
dl.acm.org | doi.org | CNR IRIS | CNR IRIS
2026
Journal article
Open Access
A classification-aware super-resolution framework for ship targets in SAR imagery
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
High-resolution imagery plays a critical role in improving the performance of visual recognition tasks such as classification, detection, and segmentation. In many domains, including remote sensing and surveillance, low-resolution images can limit the accuracy of automated analysis. To address this, super-resolution (SR) techniques have been widely adopted to attempt to reconstruct high-resolution images from low-resolution inputs. Related traditional approaches focus solely on enhancing image quality based on pixel-level metrics, leaving the relationship between super-resolved image fidelity and downstream classification performance largely underexplored. This raises a key question: can integrating classification objectives directly into the super-resolution process further improve classification accuracy? In this paper, we try to respond to this question by investigating the relationship between super-resolution and classification through the deployment of a specialised algorithmic strategy. We propose a novel methodology that increases the resolution of synthetic aperture radar imagery by optimising loss functions that account for both image quality and classification performance. Our approach improves image quality, as measured by scientifically ascertained image quality indicators, while also enhancing classification accuracy.Source: IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, vol. 19, pp. 6614-6622
DOI: 10.1109/jstars.2026.3655550
Metrics:
Awais Ch Muhammad, Reggiannini Marco, Moroni Davide, Karakus Oktay
High-resolution imagery plays a critical role in improving the performance of visual recognition tasks such as classification, detection, and segmentation. In many domains, including remote sensing and surveillance, low-resolution images can limit the accuracy of automated analysis. To address this, super-resolution (SR) techniques have been widely adopted to attempt to reconstruct high-resolution images from low-resolution inputs. Related traditional approaches focus solely on enhancing image quality based on pixel-level metrics, leaving the relationship between super-resolved image fidelity and downstream classification performance largely underexplored. This raises a key question: can integrating classification objectives directly into the super-resolution process further improve classification accuracy? In this paper, we try to respond to this question by investigating the relationship between super-resolution and classification through the deployment of a specialised algorithmic strategy. We propose a novel methodology that increases the resolution of synthetic aperture radar imagery by optimising loss functions that account for both image quality and classification performance. Our approach improves image quality, as measured by scientifically ascertained image quality indicators, while also enhancing classification accuracy.Source: IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, vol. 19, pp. 6614-6622
DOI: 10.1109/jstars.2026.3655550
Metrics:
See at:
CNR IRIS | ieeexplore.ieee.org | CNR IRIS | CNR IRIS
2026
Journal article
Open Access
Stem education and ICT-enhanced tools for students with disabilities: a five-year review
Pieriboni Giuditta, Buzzi Marina, Leporini Barbara
Education is a universal right that must include anyone. However, students with disabilities often experience significant challenges, obstacles, and frustration when interacting with digital tools and systems, particularly when performing STEM (Science, Technology, Engineering, Mathematics) education paths. This review investigates the current state of ICT-enhanced tools designed to support inclusive STEM education, focusing on the use of applications, systems, games, robots, and platforms aimed at supporting students with disabilities. While the literature reveals growing interest and numerous theoretical frameworks, there remains a lack of concrete, widely adopted applications and systems. Our results highlight the potential of Artificial Intelligence (AI), especially generative AI, in enabling personalized, accessible, and engaging learning experiences conform to the Universal Design for Learning (UDL) principles. However, the full potential of AI in inclusive education is still in its infancy and needs to be further discovered. Co-design with educators, stakeholders, and individuals with disabilities would be essential to accelerate this progress and ensure that future education technologies and systems are truly tailored and suitable for all students, regardless of their abilities and preferences. Digital tools must meet the needs of diverse populations with different disabilities to guarantee equal access to education, while engaging, motivating, and empowering students in science and technology.Source: UNIVERSAL ACCESS IN THE INFORMATION SOCIETY, vol. 25 (issue 1)
DOI: 10.1007/s10209-025-01282-8
Metrics:
Pieriboni Giuditta, Buzzi Marina, Leporini Barbara
Education is a universal right that must include anyone. However, students with disabilities often experience significant challenges, obstacles, and frustration when interacting with digital tools and systems, particularly when performing STEM (Science, Technology, Engineering, Mathematics) education paths. This review investigates the current state of ICT-enhanced tools designed to support inclusive STEM education, focusing on the use of applications, systems, games, robots, and platforms aimed at supporting students with disabilities. While the literature reveals growing interest and numerous theoretical frameworks, there remains a lack of concrete, widely adopted applications and systems. Our results highlight the potential of Artificial Intelligence (AI), especially generative AI, in enabling personalized, accessible, and engaging learning experiences conform to the Universal Design for Learning (UDL) principles. However, the full potential of AI in inclusive education is still in its infancy and needs to be further discovered. Co-design with educators, stakeholders, and individuals with disabilities would be essential to accelerate this progress and ensure that future education technologies and systems are truly tailored and suitable for all students, regardless of their abilities and preferences. Digital tools must meet the needs of diverse populations with different disabilities to guarantee equal access to education, while engaging, motivating, and empowering students in science and technology.Source: UNIVERSAL ACCESS IN THE INFORMATION SOCIETY, vol. 25 (issue 1)
DOI: 10.1007/s10209-025-01282-8
Metrics:
See at:
CNR IRIS | link.springer.com | CNR IRIS
2026
Journal article
Open Access
Robustness of complexity estimation in event-driven signals against accuracy of event detection method
Cafiso Marco, Paradisi Paolo
Complexity has gained recent attention in machine learning for its ability to extract synthetic information from large datasets. Complex dynamical systems are characterized by temporal complexity associated with intermittent birth–death events of self-organizing behavior. These rapid transition events (RTEs) can be modeled as a stochastic point process on the time axis, with inter-event times (IETs) revealing rich dynamics. In particular, IETs with power-law distribution mark a departure from the Poisson statistics and indicate the presence of nontrivial complexity that is quantified by the power-law exponent μ of the IET distribution. However, detection of RTEs in noisy signals remains a challenge, since false positives can obscure the statistical structure of the underlying process. In this paper, we address the problem of quantifying the effect of the event detection tool on the accuracy of complexity estimation. This is reached through a systematic evaluation of the Event-Driven Diffusion Scaling (EDDiS) algorithm, a tool exploiting event-driven diffusion to estimate temporal complexity. After introducing the event detection method RTE-Finder (RTEF), we assess the performance of the RTEF-EDDiS pipeline using event-driven synthetic signals. The reliability of the RTEF is found to strongly depend on parameters such as the percentile and the number of false positives can be much higher than the number of genuine complex events. Despite this, we found that the complexity estimation is quite robust with respect to the rate of false positives. In many of the studied cases, the second moment scaling H appears to even improve as the rate of false positives increases, reaching estimation errors of about 4−7%.Source: CHAOS, SOLITONS AND FRACTALS, vol. 208 (issue 118264)
DOI: 10.1016/j.chaos.2026.118264
DOI: 10.2139/ssrn.5342885
DOI: 10.48550/arxiv.2506.06168
Project(s): Future Artificial Intelligence Research - Spoke 1 “Human-centered AI”
Metrics:
Cafiso Marco, Paradisi Paolo
Complexity has gained recent attention in machine learning for its ability to extract synthetic information from large datasets. Complex dynamical systems are characterized by temporal complexity associated with intermittent birth–death events of self-organizing behavior. These rapid transition events (RTEs) can be modeled as a stochastic point process on the time axis, with inter-event times (IETs) revealing rich dynamics. In particular, IETs with power-law distribution mark a departure from the Poisson statistics and indicate the presence of nontrivial complexity that is quantified by the power-law exponent μ of the IET distribution. However, detection of RTEs in noisy signals remains a challenge, since false positives can obscure the statistical structure of the underlying process. In this paper, we address the problem of quantifying the effect of the event detection tool on the accuracy of complexity estimation. This is reached through a systematic evaluation of the Event-Driven Diffusion Scaling (EDDiS) algorithm, a tool exploiting event-driven diffusion to estimate temporal complexity. After introducing the event detection method RTE-Finder (RTEF), we assess the performance of the RTEF-EDDiS pipeline using event-driven synthetic signals. The reliability of the RTEF is found to strongly depend on parameters such as the percentile and the number of false positives can be much higher than the number of genuine complex events. Despite this, we found that the complexity estimation is quite robust with respect to the rate of false positives. In many of the studied cases, the second moment scaling H appears to even improve as the rate of false positives increases, reaching estimation errors of about 4−7%.Source: CHAOS, SOLITONS AND FRACTALS, vol. 208 (issue 118264)
DOI: 10.1016/j.chaos.2026.118264
DOI: 10.2139/ssrn.5342885
DOI: 10.48550/arxiv.2506.06168
Project(s): Future Artificial Intelligence Research - Spoke 1 “Human-centered AI”
Metrics:
See at:
arXiv.org e-Print Archive | Chaos Solitons & Fractals | CNR IRIS | www.sciencedirect.com | doi.org | doi.org | CNR IRIS
2026
Other
Open Access
Tecniche di IA Generativa per l'arricchimento di dataset di immagini
Torbidoni Giacomo, Del Corso Giulio, Papini Oscar, Tortorella Domenico
Negli ultimi anni, i sistemi di Deep Learning hanno assunto un ruolo centrale in numerosi ambiti. L'evoluzione delle architetture neurali ha portato allo sviluppo di modelli sempre più complessi e performanti, i quali tuttavia richiedono dataset di addestramento composti anche da decine di milioni di immagini In molti contesti reali non è possibile raggiungere tali numerosità, sia per i costi elevati in termini di tempo e denaro per l'acquisizione e l'annotazione, sia talvolta per la natura specifica del dominio applicativo. Negli ultimi anni, l'emergere dei modelli generativi ha aperto nuove prospettive nell'ambito della data augmentation, consentendo la produzione di dati sintetici realistici e coerenti. Questa tesi, che costituisce la relazione del tirocinio svolto da novembre 2025 a febbraio 2026 presso l'Istituto di Scienza e Tecnologie dell'Informazione del Consiglio Nazionale delle Ricerche (ISTI-CNR) di Pisa nell'ambito del progetto europeo FAITH, di cui ISTI-CNR è il leader del sottoprogetto pilota sul trasporto pubblico, si concentra sulla progettazione, implementazione e validazione di un processo di data augmentation avanzata basato su modelli di intelligenza artificiale generativa, con il duplice obiettivo di ampliare numericamente un dataset esistente e al contempo sviluppare una metodologia controllata e riproducibile per la generazione di immagini sintetiche realistiche e coerenti da integrare in contesti visivi reali.
Torbidoni Giacomo, Del Corso Giulio, Papini Oscar, Tortorella Domenico
Negli ultimi anni, i sistemi di Deep Learning hanno assunto un ruolo centrale in numerosi ambiti. L'evoluzione delle architetture neurali ha portato allo sviluppo di modelli sempre più complessi e performanti, i quali tuttavia richiedono dataset di addestramento composti anche da decine di milioni di immagini In molti contesti reali non è possibile raggiungere tali numerosità, sia per i costi elevati in termini di tempo e denaro per l'acquisizione e l'annotazione, sia talvolta per la natura specifica del dominio applicativo. Negli ultimi anni, l'emergere dei modelli generativi ha aperto nuove prospettive nell'ambito della data augmentation, consentendo la produzione di dati sintetici realistici e coerenti. Questa tesi, che costituisce la relazione del tirocinio svolto da novembre 2025 a febbraio 2026 presso l'Istituto di Scienza e Tecnologie dell'Informazione del Consiglio Nazionale delle Ricerche (ISTI-CNR) di Pisa nell'ambito del progetto europeo FAITH, di cui ISTI-CNR è il leader del sottoprogetto pilota sul trasporto pubblico, si concentra sulla progettazione, implementazione e validazione di un processo di data augmentation avanzata basato su modelli di intelligenza artificiale generativa, con il duplice obiettivo di ampliare numericamente un dataset esistente e al contempo sviluppare una metodologia controllata e riproducibile per la generazione di immagini sintetiche realistiche e coerenti da integrare in contesti visivi reali.
2026
Other
Open Access
Robust computer vision for POCUS: achieving reliability in ultrasound imaging
Ignesti Giacomo
The rapid proliferation of Deep Learning (DL) has fundamentally transformed the scientific landscape, shifting the paradigm of problem-solving from explicit, algorithmic logic to data-driven feature learning. However, as these technologies permeate high-stakes domains such as medicine and precision agriculture, a critical tension has emerged. The very non-linearity that grants Neural Networks their immense expressive power, as described by the Universal Approximation Theorem, simultaneously causes intrinsic opacity, gradient instability, and a susceptibility to spurious correlations. Consequently, the deployment of these black-box models faces significant hurdles in terms of reliability and trust, necessitating the development of Artificial Intelligence (AI) that is not merely accurate but also fundamentally reliable and interpretable. The Clinical Context: Telemedicine and the TiAssisto Project. This thesis grounds its research in the critical context of the SARS-CoV-2 pandemic, which exposed systemic fragilities in healthcare and highlighted the need for resilient, decentralised care models. A significant portion of the applied research was conducted within the TiAssisto project, a server-based web platform designed to support multi-pathological patients by bridging the gap between hospitalisation and home care. To operationalise the vision of Medicine 5.0—a human-centric approach to healthcare—TiAssisto integrates a hybrid AI framework. This comprises a Strong AI component utilising a rule-based algorithm and logic for monitoring vital signs, and a Weak AI component employing a DL Model for the automatic classification of medical imaging. The platform prioritises patient security through strict local encryption and anonymisation protocols, anticipating the growing necessity for Reliable AI architectures. The Challenge of Point-of-Care Ultrasound (POCUS). Within this framework, the thesis identifies Lung Ultrasound (LUS) as a pivotal technology for telemedicine, given its portability and safety. However, LUS presents unique challenges for Computer Vision; unlike MRI or CT scans, LUS images are characterised by a low signal-to-noise ratio and are artefact-rich. Diagnosis often relies on interpreting acoustic artefacts, such as "A-lines" (horizontal repetitions indicating a healthy lung) or "B-lines" (vertical comet-tail artefacts indicating pathology), rather than direct anatomical visualization. To address the computational constraints of Point-of-Care (POCUS) devices and the scarcity of labelled data, the research proposes a novel Efficient Adaptive Ensembling methodology. This approach utilises the EfficientNet-b0 DL architecture as a backbone due to its optimal balance of accuracy and complexity. By training weak learners on data subsets to induce diversity and subsequently synthesising their deep features via a trainable linear combination layer, the model achieves robustness without the computational cost of traditional ensembles. When validated on a large public LUS dataset, this ensemble model achieved 100% accuracy in distinguishing between COVID-19, Pneumonia, and Healthy classes, significantly outperforming previous state-of-the-art models. Crucially, it maintained high efficiency, requiring only 0.78 G-FLOPs, making it suitable for real-time deployment on edge devices. Foundations of Reliable AI: From Diagnosis to Constraint. While achieving high accuracy is essential, the thesis argues that clinical adoption hinges on trust. Initial efforts utilised post-hoc Explainable AI (XAI) methods, specifically Grad-CAM, to visually validate that the LUS model focused on clinically relevant areas, such as the pleural line. However, the research identifies a critical limitation in post-hoc analysis: it is purely diagnostic and cannot correct a model’s flawed reasoning. To transcend this limitation, the thesis introduces a paradigm shift towards Intrinsically Guided Training. The research proposes moving the generation of Class Activation Maps (CAM) into the training loop itself. By incorporating a custom Contrastive Loss term, the training process actively penalises the model if its internal attention does not align with the relevant input features. This concept evolved into the Batch-CAM paradigm. This advanced methodology replaces instance-specific reconstruction with a Class Prototype approach, where the model compares batch-averaged CAMs against a learned prototype of the target class. This enforces semantic consistency, compelling the model to learn a generalised representation of a class (e.g., the concept of a "T-shirt" or a specific digit) rather than memorising individual examples. Experimental validation on MNIST and Fashion-MNIST demonstrated that this approach yields models with competitive accuracy and significantly superior, human-intelligible internal representations. Conclusion and Future Outlook. The thesis concludes that the path to Reliable AI lies in shifting from passive analysis to active constraint during the learning process. By embedding interpretability into the loss function, models can be guided to learn for the right reasons, mitigating the risk of spurious correlations. The first application of the method to LUS interpretation yields high accuracy, and the systems, now partially guided, seem to learn meaningful anatomical details to inform their decisions. Looking forward, the research outlines a trajectory towards Continual Learning to prevent catastrophic forgetting as medical protocols evolve, and Federated Learning to facilitate privacy-preserving training on decentralised data. Ultimately, integrating these Plausible AI models into hybrid Decision Support Systems offers a structure for the future of healthcare systems that are efficient, transparent, and trustworthy partners in human decision-making.
Ignesti Giacomo
The rapid proliferation of Deep Learning (DL) has fundamentally transformed the scientific landscape, shifting the paradigm of problem-solving from explicit, algorithmic logic to data-driven feature learning. However, as these technologies permeate high-stakes domains such as medicine and precision agriculture, a critical tension has emerged. The very non-linearity that grants Neural Networks their immense expressive power, as described by the Universal Approximation Theorem, simultaneously causes intrinsic opacity, gradient instability, and a susceptibility to spurious correlations. Consequently, the deployment of these black-box models faces significant hurdles in terms of reliability and trust, necessitating the development of Artificial Intelligence (AI) that is not merely accurate but also fundamentally reliable and interpretable. The Clinical Context: Telemedicine and the TiAssisto Project. This thesis grounds its research in the critical context of the SARS-CoV-2 pandemic, which exposed systemic fragilities in healthcare and highlighted the need for resilient, decentralised care models. A significant portion of the applied research was conducted within the TiAssisto project, a server-based web platform designed to support multi-pathological patients by bridging the gap between hospitalisation and home care. To operationalise the vision of Medicine 5.0—a human-centric approach to healthcare—TiAssisto integrates a hybrid AI framework. This comprises a Strong AI component utilising a rule-based algorithm and logic for monitoring vital signs, and a Weak AI component employing a DL Model for the automatic classification of medical imaging. The platform prioritises patient security through strict local encryption and anonymisation protocols, anticipating the growing necessity for Reliable AI architectures. The Challenge of Point-of-Care Ultrasound (POCUS). Within this framework, the thesis identifies Lung Ultrasound (LUS) as a pivotal technology for telemedicine, given its portability and safety. However, LUS presents unique challenges for Computer Vision; unlike MRI or CT scans, LUS images are characterised by a low signal-to-noise ratio and are artefact-rich. Diagnosis often relies on interpreting acoustic artefacts, such as "A-lines" (horizontal repetitions indicating a healthy lung) or "B-lines" (vertical comet-tail artefacts indicating pathology), rather than direct anatomical visualization. To address the computational constraints of Point-of-Care (POCUS) devices and the scarcity of labelled data, the research proposes a novel Efficient Adaptive Ensembling methodology. This approach utilises the EfficientNet-b0 DL architecture as a backbone due to its optimal balance of accuracy and complexity. By training weak learners on data subsets to induce diversity and subsequently synthesising their deep features via a trainable linear combination layer, the model achieves robustness without the computational cost of traditional ensembles. When validated on a large public LUS dataset, this ensemble model achieved 100% accuracy in distinguishing between COVID-19, Pneumonia, and Healthy classes, significantly outperforming previous state-of-the-art models. Crucially, it maintained high efficiency, requiring only 0.78 G-FLOPs, making it suitable for real-time deployment on edge devices. Foundations of Reliable AI: From Diagnosis to Constraint. While achieving high accuracy is essential, the thesis argues that clinical adoption hinges on trust. Initial efforts utilised post-hoc Explainable AI (XAI) methods, specifically Grad-CAM, to visually validate that the LUS model focused on clinically relevant areas, such as the pleural line. However, the research identifies a critical limitation in post-hoc analysis: it is purely diagnostic and cannot correct a model’s flawed reasoning. To transcend this limitation, the thesis introduces a paradigm shift towards Intrinsically Guided Training. The research proposes moving the generation of Class Activation Maps (CAM) into the training loop itself. By incorporating a custom Contrastive Loss term, the training process actively penalises the model if its internal attention does not align with the relevant input features. This concept evolved into the Batch-CAM paradigm. This advanced methodology replaces instance-specific reconstruction with a Class Prototype approach, where the model compares batch-averaged CAMs against a learned prototype of the target class. This enforces semantic consistency, compelling the model to learn a generalised representation of a class (e.g., the concept of a "T-shirt" or a specific digit) rather than memorising individual examples. Experimental validation on MNIST and Fashion-MNIST demonstrated that this approach yields models with competitive accuracy and significantly superior, human-intelligible internal representations. Conclusion and Future Outlook. The thesis concludes that the path to Reliable AI lies in shifting from passive analysis to active constraint during the learning process. By embedding interpretability into the loss function, models can be guided to learn for the right reasons, mitigating the risk of spurious correlations. The first application of the method to LUS interpretation yields high accuracy, and the systems, now partially guided, seem to learn meaningful anatomical details to inform their decisions. Looking forward, the research outlines a trajectory towards Continual Learning to prevent catastrophic forgetting as medical protocols evolve, and Federated Learning to facilitate privacy-preserving training on decentralised data. Ultimately, integrating these Plausible AI models into hybrid Decision Support Systems offers a structure for the future of healthcare systems that are efficient, transparent, and trustworthy partners in human decision-making.
2026
Other
Restricted
AI4GO: AI Framework for Gamma-Oryzanol-rich oil
Martinelli Massimo
A Collaborative Framework for Non-Invasive Monitoring and Predictive Quality Control and Zero-Waste Efficiency in Rice Bran Oil.
Martinelli Massimo
A Collaborative Framework for Non-Invasive Monitoring and Predictive Quality Control and Zero-Waste Efficiency in Rice Bran Oil.
2026
Journal article
Open Access
Un metodo per l'analisi di strutture oniriche nelle opere d'arte. Applicazione a una canzone di Battiato
Magrini Massimo, Barcaro Umberto
Viene proposto un metodo per l’analisi delle opere d’arte. Esso si basa sul riconoscimento e sulla descrizione di ca-ratteristiche che presentano una forte somiglianza con le strutture che caratterizzano l’esperienza onirica. I modelli onirici presi in esame sono quelli identificati dalla ricerca sperimentale sui sogni. Il metodo trae ispirazione dall’idea che una stretta connessione strutturale tra le opere d’arte e l’esperienza onirica possa instaurare una comunicazione emotiva inconscia tra il creatore e il destinatario dell’opera d’arte. Viene presentato un esempio di applicazione del metodo: si tratta di una canzone di Franco Battiato pubblicata nel 1980. Sebbene venga riportata una singola appli-cazione, il metodo sembra essere generalmente applicabile alle opere d’arte.Source: ITINERA, vol. 31, pp. 208-220
DOI: 10.54103/2039-9251/32303
Metrics:
Magrini Massimo, Barcaro Umberto
Viene proposto un metodo per l’analisi delle opere d’arte. Esso si basa sul riconoscimento e sulla descrizione di ca-ratteristiche che presentano una forte somiglianza con le strutture che caratterizzano l’esperienza onirica. I modelli onirici presi in esame sono quelli identificati dalla ricerca sperimentale sui sogni. Il metodo trae ispirazione dall’idea che una stretta connessione strutturale tra le opere d’arte e l’esperienza onirica possa instaurare una comunicazione emotiva inconscia tra il creatore e il destinatario dell’opera d’arte. Viene presentato un esempio di applicazione del metodo: si tratta di una canzone di Franco Battiato pubblicata nel 1980. Sebbene venga riportata una singola appli-cazione, il metodo sembra essere generalmente applicabile alle opere d’arte.Source: ITINERA, vol. 31, pp. 208-220
DOI: 10.54103/2039-9251/32303
Metrics:
See at:
CNR IRIS | riviste.unimi.it | CNR IRIS
2026
Journal article
Open Access
A study of the sources of nap mentation in the light of the continuity theory
Curzio Olivia, Di Galante Marco, Faraguna Ugo, Maestro Sandra, Magrini Massimo, Paradisi Paolo, Sebastiani Laura
Summary. The aim of this paper was to describe the properties of sources of mentation during morning naps. The analyzed sources, which consisted in waking-life events and concerns, were obtained by asking the participants, immediately after their naps, for a report of their mentation and for associations with the various mentation items. The participants comprised 18 Italian students aged 19-22 years, 12 females and 6 males. All participants reported mentation, which was rich and articulated in most cases, and all but one indicated the mentation sources with precision and in detail. The properties of these sources were consistent with those reported in the literature for night dreams. In fact: (a) most participants (n = 14) indicated three or more detailed sources; (b) all of the sources were correlated with present concerns of the participants; (c) most participants (n=13) indicated more than one present concern; (d) in all cases conceptual and/ or emotional links connecting at least two sources (three or more for nine participants) provided the set of sources with a network structure; (e) for almost all the participants (n=16) the links between sources determined a change of negative waking-life content into positive, and this change was actually realized in the mentation experience.Source: INTERNATIONAL JOURNAL OF DREAM RESEARCH, vol. 19 (issue 1), pp. 133-139
DOI: 10.11588/ijodr.2026.1.113629
Metrics:
Curzio Olivia, Di Galante Marco, Faraguna Ugo, Maestro Sandra, Magrini Massimo, Paradisi Paolo, Sebastiani Laura
Summary. The aim of this paper was to describe the properties of sources of mentation during morning naps. The analyzed sources, which consisted in waking-life events and concerns, were obtained by asking the participants, immediately after their naps, for a report of their mentation and for associations with the various mentation items. The participants comprised 18 Italian students aged 19-22 years, 12 females and 6 males. All participants reported mentation, which was rich and articulated in most cases, and all but one indicated the mentation sources with precision and in detail. The properties of these sources were consistent with those reported in the literature for night dreams. In fact: (a) most participants (n = 14) indicated three or more detailed sources; (b) all of the sources were correlated with present concerns of the participants; (c) most participants (n=13) indicated more than one present concern; (d) in all cases conceptual and/ or emotional links connecting at least two sources (three or more for nine participants) provided the set of sources with a network structure; (e) for almost all the participants (n=16) the links between sources determined a change of negative waking-life content into positive, and this change was actually realized in the mentation experience.Source: INTERNATIONAL JOURNAL OF DREAM RESEARCH, vol. 19 (issue 1), pp. 133-139
DOI: 10.11588/ijodr.2026.1.113629
Metrics:
See at:
CNR IRIS | journals.ub.uni-heidelberg.de | doi.org | CNR IRIS
2026
Journal article
Open Access
An AHP experiment with the smart readiness indicator applied in a local reality
Salerno Emanuele
This paper describes an experiment intended to evaluate the appropriateness of the weighting scheme used to compute the Smart Readiness Indicator (SRI), an index introduced by the European Commission and Parliament with the double aim to measure the “intelligence” of a building and to encourage the adoption of smart technologies in both new buildings and existing constructions to be retrofitted. This index is based on a hierarchical scheme including criteria related to energy savings and flexibility as well as to the comfort, convenience and wellbeing of the building occupants. In particular, the weights to assign to the criteria related to the needs of the occupants are largely subjective, and the technical support group established by the European Commission to draw up the SRI scheme did not find any scientific evidence to differentiate them, so all of them received equal weights. Exploiting Saaty’s Analysis of the Hierarchical Process (AHP), we involved a local group of engineers, architects, technicians and administrators to obtain cardinal rankings for these criteria. The results are likely to depend on the local context where our inquiry has been conducted, central Italy, but also show that the opinions of our experts are diverse, even within professionally homogeneous subgroups, and do not justify the equal-weight choice made by the technical support group. Our results can also constitute one of the many expected local contributions to the revision of the present SRI scheme.Source: ENERGY AND BUILDINGS, vol. 370
DOI: 10.1016/j.enbuild.2026.118052
Project(s): Green, Utility, Accessibility, Resilience, Digital Integration, Ability, Nature, Sustainability
Metrics:
Salerno Emanuele
This paper describes an experiment intended to evaluate the appropriateness of the weighting scheme used to compute the Smart Readiness Indicator (SRI), an index introduced by the European Commission and Parliament with the double aim to measure the “intelligence” of a building and to encourage the adoption of smart technologies in both new buildings and existing constructions to be retrofitted. This index is based on a hierarchical scheme including criteria related to energy savings and flexibility as well as to the comfort, convenience and wellbeing of the building occupants. In particular, the weights to assign to the criteria related to the needs of the occupants are largely subjective, and the technical support group established by the European Commission to draw up the SRI scheme did not find any scientific evidence to differentiate them, so all of them received equal weights. Exploiting Saaty’s Analysis of the Hierarchical Process (AHP), we involved a local group of engineers, architects, technicians and administrators to obtain cardinal rankings for these criteria. The results are likely to depend on the local context where our inquiry has been conducted, central Italy, but also show that the opinions of our experts are diverse, even within professionally homogeneous subgroups, and do not justify the equal-weight choice made by the technical support group. Our results can also constitute one of the many expected local contributions to the revision of the present SRI scheme.Source: ENERGY AND BUILDINGS, vol. 370
DOI: 10.1016/j.enbuild.2026.118052
Project(s): Green, Utility, Accessibility, Resilience, Digital Integration, Ability, Nature, Sustainability
Metrics:
See at:
CNR IRIS | www.sciencedirect.com | CNR IRIS