Typology
operator:
and / or
Language
operator:
and / or
Date
operator:
and / or
Rights
operator:
and / or
2024
Other
Open Access
Stretch your tentacles, POLPO-net: a POLymorphic PrObabilistic approach to greedily approximate model uncertainties
Del Corso G., Caudai C., Kuruoglu E. E., Colantonio S.
A POLymorphic PrObabilistic approach to greedily approximate uncertainties avoiding re-training of costly deep neural networks.DOI: 10.5281/zenodo.12780350
Metrics:
Del Corso G., Caudai C., Kuruoglu E. E., Colantonio S.
A POLymorphic PrObabilistic approach to greedily approximate uncertainties avoiding re-training of costly deep neural networks.DOI: 10.5281/zenodo.12780350
Metrics:
2024
Other
Open Access
EchoLocator: an open source Python package for the standardisation of echographic images in multicentre analysis
Del Corso G., De Rosa L., Pascali M. A., Faita F., Colantonio S.
In this technical report, we provide a fully automated preprocessing package, developed entirely in Python 3.6, to reduce such heterogeneity in US images. This package allows the automatic removal of echographic watermarks, cropping and centering the echographic cone. Moreover, the echographic cone is converted in a rectangular region. The EchoLocator package is freely available on GitHub.DOI: 10.32079/isti-tr-2024/003
DOI: https://doi.org/10.32079/isti-tr-2024/003
Metrics:
Del Corso G., De Rosa L., Pascali M. A., Faita F., Colantonio S.
In this technical report, we provide a fully automated preprocessing package, developed entirely in Python 3.6, to reduce such heterogeneity in US images. This package allows the automatic removal of echographic watermarks, cropping and centering the echographic cone. Moreover, the echographic cone is converted in a rectangular region. The EchoLocator package is freely available on GitHub.DOI: 10.32079/isti-tr-2024/003
DOI: https://doi.org/10.32079/isti-tr-2024/003
Metrics:
2024
Other
Open Access
NA DAtabase: generator of probabilistic synthetic geometrical shape dataset
Volpini F., Caudai C., Del Corso G., Colantonio S.
In this technical report, we detail NA DA (Not-A-DAtabase), an open-source software writ- ten in Python that generates datasets of regular two-dimensional geometric shapes based on probabilistic distributions (https://github.com/GDelCorso/NA DAtabase.git). NA DA comes with an intuitive GUI (Graphical User Interface) that allows users to define shapes, colors, and distributions of features of datasets consisting of image sets and CSV files containing metadata for each element. These databases can be saved to provide a unique identifier of the dataset, allowing perfect reproducibility or easy modification of the dataset using the GUI or directly by calling the generator class. Therefore, NA DA is a tool to help and support the investigation of trustworthiness, overconfidence, uncertainty, and computation time of machine learning and deep learning models.DOI: 10.32079/isti-tr-2024/004
Metrics:
Volpini F., Caudai C., Del Corso G., Colantonio S.
In this technical report, we detail NA DA (Not-A-DAtabase), an open-source software writ- ten in Python that generates datasets of regular two-dimensional geometric shapes based on probabilistic distributions (https://github.com/GDelCorso/NA DAtabase.git). NA DA comes with an intuitive GUI (Graphical User Interface) that allows users to define shapes, colors, and distributions of features of datasets consisting of image sets and CSV files containing metadata for each element. These databases can be saved to provide a unique identifier of the dataset, allowing perfect reproducibility or easy modification of the dataset using the GUI or directly by calling the generator class. Therefore, NA DA is a tool to help and support the investigation of trustworthiness, overconfidence, uncertainty, and computation time of machine learning and deep learning models.DOI: 10.32079/isti-tr-2024/004
Metrics:
2024
Journal article
Open Access
An original systematic review: use of artificial intelligence and unmet needs in eosinophilic oesophagitis from COVID-19 era
Bardelli S., Del Corso G., Ciantelli M., Del Pistoia M., Scaramuzzo R. T., Cuttano A.
In this work, we investigated the effectiveness of a digital game-based learning (DGBL) methodology for remote training using DGBL software specifically designed for neonatal resuscitation. The DGBL approach was validated using state-of-the-art statistical methods in a cohort of 52 anesthesiology trainees and compared to a homogeneous retrospective control group of pediatric trainees with comparable prior knowledge, who underwent an in-person training course using the same digital serious game. Scores obtained during each game session were recorded using in-house software and used to assess progress in flowchart knowledge, decision-making time, timing of assisted ventilation, and ability to check equipment. The results confirmed that the DGBL-based remote training approach is a valuable tool that offers an interactive, effective, and engaging learning experience. Future developments will integrate an adaptive AI-based agent to further enhance the game's effectiveness.Source: GAMES FOR HEALTH JOURNAL, vol. 13 (issue 6), pp. 452-458
DOI: 10.1089/g4h.2023.0197
Metrics:
Bardelli S., Del Corso G., Ciantelli M., Del Pistoia M., Scaramuzzo R. T., Cuttano A.
In this work, we investigated the effectiveness of a digital game-based learning (DGBL) methodology for remote training using DGBL software specifically designed for neonatal resuscitation. The DGBL approach was validated using state-of-the-art statistical methods in a cohort of 52 anesthesiology trainees and compared to a homogeneous retrospective control group of pediatric trainees with comparable prior knowledge, who underwent an in-person training course using the same digital serious game. Scores obtained during each game session were recorded using in-house software and used to assess progress in flowchart knowledge, decision-making time, timing of assisted ventilation, and ability to check equipment. The results confirmed that the DGBL-based remote training approach is a valuable tool that offers an interactive, effective, and engaging learning experience. Future developments will integrate an adaptive AI-based agent to further enhance the game's effectiveness.Source: GAMES FOR HEALTH JOURNAL, vol. 13 (issue 6), pp. 452-458
DOI: 10.1089/g4h.2023.0197
Metrics:
See at:
CNR IRIS 
| www.liebertpub.com | Games for Health Journal 
| CNR IRIS | CNR IRIS | Games for Health Journal
2024
Conference article
Restricted
Radiomics-based reliable predictions of side effects after radiotherapy for prostate cancer
Del Corso G., Pachetti E., Buongiorno R., Rodrigues A. C., Germanese D., Pascali M. A., Almeida J., Rodrigues N., Tsiknakis M., Papanikolaou N., Regge D., Marias K., Consortium Procancer-I, Colantonio S.
This work offers insight into the effectiveness of probabilistic models, specifically those based on ensemble approximations, in predicting adverse side effects following radiotherapy for prostate cancer. We trained a random forest model on radiomic features from 134 T2-weighted Magnetic Resonance (MRI) images of the prostate gland to identify patients experiencing acute or chronic rectal and urinary toxicity (AU-ROC ranging from 61.4% for endorectal coil acquisitions to 70.8% for the full dataset). We evaluated the reliability of the predictions using an ensemble approximation of simplified random forests obtained by an adaptive procedure of random subsampling of the training data. We used this reliability score to define a not-confident class and then recompute performance metrics more in accordance with a probabilistic approach. The outcomes we obtained (up to 7.9% increase in accuracy) indicate the approximated probabilistic models pledge more reliable predictions, thus being suitable for further investigation.DOI: 10.1109/isbi56570.2024.10635233
Project(s): ProCAncer-I
Metrics:
Del Corso G., Pachetti E., Buongiorno R., Rodrigues A. C., Germanese D., Pascali M. A., Almeida J., Rodrigues N., Tsiknakis M., Papanikolaou N., Regge D., Marias K., Consortium Procancer-I, Colantonio S.
This work offers insight into the effectiveness of probabilistic models, specifically those based on ensemble approximations, in predicting adverse side effects following radiotherapy for prostate cancer. We trained a random forest model on radiomic features from 134 T2-weighted Magnetic Resonance (MRI) images of the prostate gland to identify patients experiencing acute or chronic rectal and urinary toxicity (AU-ROC ranging from 61.4% for endorectal coil acquisitions to 70.8% for the full dataset). We evaluated the reliability of the predictions using an ensemble approximation of simplified random forests obtained by an adaptive procedure of random subsampling of the training data. We used this reliability score to define a not-confident class and then recompute performance metrics more in accordance with a probabilistic approach. The outcomes we obtained (up to 7.9% increase in accuracy) indicate the approximated probabilistic models pledge more reliable predictions, thus being suitable for further investigation.DOI: 10.1109/isbi56570.2024.10635233
Project(s): ProCAncer-I
Metrics:
2024
Conference article
Open Access
From Covid-19 detection to cancer grading: how medical-AI is boosting clinical diagnostics and may improve treatment
Berti A., Buongiorno R., Carloni G., Caudai C., Conti F., Del Corso G., Germanese D., Moroni D., Pachetti E., Pascali M. A., Colantonio S.
The integration of artificial intelligence (AI) into medical imaging has guided an era of transformation in healthcare. This paper presents the research activities that a multidisciplinary research group within the Signals and Images Lab of the Institute of Information Science and Technologies of the National Research Council of Italy is carrying out to explore the great potential of AI in medical imaging. From the convolutional neural network-based segmentation of Covid-19 lung patterns to the radiomic signature for benign/malignant breast nodule discrimination, to the automatic grading of prostate cancer, this work highlights the paradigm shift that AI has brought to medical imaging, revolutionizing diagnosis and patient care.Source: CEUR WORKSHOP PROCEEDINGS, vol. 3762, pp. 336-341. Naples, Italy, 29-30/05/2024
Berti A., Buongiorno R., Carloni G., Caudai C., Conti F., Del Corso G., Germanese D., Moroni D., Pachetti E., Pascali M. A., Colantonio S.
The integration of artificial intelligence (AI) into medical imaging has guided an era of transformation in healthcare. This paper presents the research activities that a multidisciplinary research group within the Signals and Images Lab of the Institute of Information Science and Technologies of the National Research Council of Italy is carrying out to explore the great potential of AI in medical imaging. From the convolutional neural network-based segmentation of Covid-19 lung patterns to the radiomic signature for benign/malignant breast nodule discrimination, to the automatic grading of prostate cancer, this work highlights the paradigm shift that AI has brought to medical imaging, revolutionizing diagnosis and patient care.Source: CEUR WORKSHOP PROCEEDINGS, vol. 3762, pp. 336-341. Naples, Italy, 29-30/05/2024
See at:
ceur-ws.org | CNR IRIS | CNR IRIS
2024
Journal article
Open Access
Few-shot conditional learning: automatic and reliable device classification for medical test equipment
Pachetti E., Del Corso G., Bardelli S., Colantonio S.
: The limited availability of specialized image databases (particularly in hospitals, where tools vary between providers) makes it difficult to train deep learning models. This paper presents a few-shot learning methodology that uses a pre-trained ResNet integrated with an encoder as a backbone to encode conditional shape information for the classification of neonatal resuscitation equipment from less than 100 natural images. The model is also strengthened by incorporating a reliability score, which enriches the prediction with an estimation of classification reliability. The model, whose performance is cross-validated, reached a median accuracy performance of over 99% (and a lower limit of 73.4% for the least accurate model/fold) using only 87 meta-training images. During the test phase on complex natural images, performance was slightly degraded due to a sub-optimal segmentation strategy (FastSAM) required to maintain the real-time inference phase (median accuracy 87.25%). This methodology proves to be excellent for applying complex classification models to contexts (such as neonatal resuscitation) that are not available in public databases. Improvements to the automatic segmentation strategy prior to the extraction of conditional information will allow a natural application in simulation and hospital settings.Source: JOURNAL OF IMAGING, vol. 10 (issue 7)
DOI: 10.3390/jimaging10070167
Metrics:
Pachetti E., Del Corso G., Bardelli S., Colantonio S.
: The limited availability of specialized image databases (particularly in hospitals, where tools vary between providers) makes it difficult to train deep learning models. This paper presents a few-shot learning methodology that uses a pre-trained ResNet integrated with an encoder as a backbone to encode conditional shape information for the classification of neonatal resuscitation equipment from less than 100 natural images. The model is also strengthened by incorporating a reliability score, which enriches the prediction with an estimation of classification reliability. The model, whose performance is cross-validated, reached a median accuracy performance of over 99% (and a lower limit of 73.4% for the least accurate model/fold) using only 87 meta-training images. During the test phase on complex natural images, performance was slightly degraded due to a sub-optimal segmentation strategy (FastSAM) required to maintain the real-time inference phase (median accuracy 87.25%). This methodology proves to be excellent for applying complex classification models to contexts (such as neonatal resuscitation) that are not available in public databases. Improvements to the automatic segmentation strategy prior to the extraction of conditional information will allow a natural application in simulation and hospital settings.Source: JOURNAL OF IMAGING, vol. 10 (issue 7)
DOI: 10.3390/jimaging10070167
Metrics:
See at:
Journal of Imaging | Journal of Imaging | IRIS Cnr | IRIS Cnr | CNR IRIS
2024
Journal article
Open Access
Adaptive behaviors, esophageal anxiety and hypervigilance modify the association between dysphagia perception and histological disease activity in eosinophilic esophagitis
Visaggi P., Del Corso G., Solinas I., Ovidi F., Adamo G., Dulmin I., Baiano Svizzero F., Bellini M., Savarino E. V., De Bortoli N.
Background aims: Esophageal symptom-specific anxiety, hypervigilance, and adaptive behaviors at mealtime may affect dysphagia reporting in patients with eosinophilic esophagitis (EoE) but this has not been investigated. Moreover, the relationship between such confounding factors and histological disease activity (HDA) is unclear. Methods: This was a prospective study on adults with EoE. Dysphagia, anxiety, and hypervigilance were assessed using specific questionnaires (i.e., mDSQ, DSS, and EHAS). Adaptive behaviors were assessed using the Pisa EoE Adaptation Questionnaire. Appropriate statistics was used to investigate correlation between dysphagia, anxiety, hypervigilance, adaptive behaviors and HDA. Results: Ninety-five patients were included. Esophageal anxiety, hypervigilance, and use of adaptive behaviors were found in about 50% of EoE patients. Esophageal anxiety and hypervigilance were significantly higher (p=0.03 for both) and adaptive behaviors were significantly more prevalent in histologically active EoE compared to EoE in remission (76.8% vs 25.6%, p<0.001). As a standalone measurement, mDSQ and DSS had AUROC of 77.7% and 75.3% for predicting HDA. Adjustments of mDSQ and DSS based on individual EHAS scores and adaptive behaviors at mealtime significantly improved the AUROC of mDSQ and DSS to 86.6% and 84.3%, respectively (p<0.05 for both). Conclusions: Higher esophageal anxiety, hypervigilance and use of adaptive behaviors are associated with active EoE and represent clinical markers of HDA. Adaptive behaviours provide complementary clinical information that is not detected by symptoms alone. The assessment of anxiety, hypervigilance, and adaptive behaviors improves the correlation between clinical and HDA in EoE.Source: THE AMERICAN JOURNAL OF GASTROENTEROLOGY
DOI: 10.14309/ajg.0000000000003272
Metrics:
Visaggi P., Del Corso G., Solinas I., Ovidi F., Adamo G., Dulmin I., Baiano Svizzero F., Bellini M., Savarino E. V., De Bortoli N.
Background aims: Esophageal symptom-specific anxiety, hypervigilance, and adaptive behaviors at mealtime may affect dysphagia reporting in patients with eosinophilic esophagitis (EoE) but this has not been investigated. Moreover, the relationship between such confounding factors and histological disease activity (HDA) is unclear. Methods: This was a prospective study on adults with EoE. Dysphagia, anxiety, and hypervigilance were assessed using specific questionnaires (i.e., mDSQ, DSS, and EHAS). Adaptive behaviors were assessed using the Pisa EoE Adaptation Questionnaire. Appropriate statistics was used to investigate correlation between dysphagia, anxiety, hypervigilance, adaptive behaviors and HDA. Results: Ninety-five patients were included. Esophageal anxiety, hypervigilance, and use of adaptive behaviors were found in about 50% of EoE patients. Esophageal anxiety and hypervigilance were significantly higher (p=0.03 for both) and adaptive behaviors were significantly more prevalent in histologically active EoE compared to EoE in remission (76.8% vs 25.6%, p<0.001). As a standalone measurement, mDSQ and DSS had AUROC of 77.7% and 75.3% for predicting HDA. Adjustments of mDSQ and DSS based on individual EHAS scores and adaptive behaviors at mealtime significantly improved the AUROC of mDSQ and DSS to 86.6% and 84.3%, respectively (p<0.05 for both). Conclusions: Higher esophageal anxiety, hypervigilance and use of adaptive behaviors are associated with active EoE and represent clinical markers of HDA. Adaptive behaviours provide complementary clinical information that is not detected by symptoms alone. The assessment of anxiety, hypervigilance, and adaptive behaviors improves the correlation between clinical and HDA in EoE.Source: THE AMERICAN JOURNAL OF GASTROENTEROLOGY
DOI: 10.14309/ajg.0000000000003272
Metrics:
See at:
CNR IRIS | journals.lww.com | The American Journal of Gastroenterology | CNR IRIS
2024
Conference article
Open Access
Facial landmark identification and data preparation can significantly improve the extraction of newborns' facial features
Del Corso G., Germanese D., Pascali M. A., Bardelli S., Cuttano A., Festante F., Guzzetta A., Rocchitelli L., Colantonio S.
Automatic extraction of facial feature can provide valuable information on the health of newborns. However, determining an optimal facial features extraction strategy, especially for preterm infants, is a challenging task due to significant differences in facial morphology and frequent pose changes. In this work, we collected video data from 10 newborns (8 preterm, 2 at term, ≤ 4 weeks post term equivalent age), obtaining a novel dataset of over 41, 000 labeled frames (Open Mouth, Closed Mouth, Tongue Protrusion). On the collected images, we applied a strong data preparation procedure (including mouth localization, cropping, and reorientation with models trained on adults), an adaptive image normalization strategy, and a proper data augmentation scheme. Thus, we trained a convolutional classifier with a large number of trainable parameters (i.e., ~1.2 million), coupled with multiple criteria to avoid overspecialization and consequent loss of generalization capability. This approach allows for highly reliable results (accuracy, precision, and recall over 92% on unseen data) and generalizes well to newborns with significantly different characteristics, even without including time-dependent information in the analysis. Therefore, these results prove that proper data preparation can narrow the gap between the classification of neonatal and adult facial features, allowing the integration of methods originally developed for adults into the complex setting of preterm infant analysis.DOI: 10.1109/fg59268.2024.10581971
Metrics:
Del Corso G., Germanese D., Pascali M. A., Bardelli S., Cuttano A., Festante F., Guzzetta A., Rocchitelli L., Colantonio S.
Automatic extraction of facial feature can provide valuable information on the health of newborns. However, determining an optimal facial features extraction strategy, especially for preterm infants, is a challenging task due to significant differences in facial morphology and frequent pose changes. In this work, we collected video data from 10 newborns (8 preterm, 2 at term, ≤ 4 weeks post term equivalent age), obtaining a novel dataset of over 41, 000 labeled frames (Open Mouth, Closed Mouth, Tongue Protrusion). On the collected images, we applied a strong data preparation procedure (including mouth localization, cropping, and reorientation with models trained on adults), an adaptive image normalization strategy, and a proper data augmentation scheme. Thus, we trained a convolutional classifier with a large number of trainable parameters (i.e., ~1.2 million), coupled with multiple criteria to avoid overspecialization and consequent loss of generalization capability. This approach allows for highly reliable results (accuracy, precision, and recall over 92% on unseen data) and generalizes well to newborns with significantly different characteristics, even without including time-dependent information in the analysis. Therefore, these results prove that proper data preparation can narrow the gap between the classification of neonatal and adult facial features, allowing the integration of methods originally developed for adults into the complex setting of preterm infant analysis.DOI: 10.1109/fg59268.2024.10581971
Metrics:
See at:
IRIS Cnr | IRIS Cnr | IRIS Cnr | doi.org | CNR IRIS | CNR IRIS
2024
Journal article
Open Access
ANN uncertainty estimates in assessing fatty liver content from ultrasound data
G. Del Corso, M. A. Pascali, C. Caudai, L. De Rosa, A. Salvati, M. Mancini, L. Ghiadoni, F. Bonino, M. R. Brunetto, S. Colantonio, F. Faita
Background and objective This article uses three different probabilistic convolutional architectures applied to ultrasound image analysis for grading Fatty Liver Content (FLC) in Metabolic Dysfunction Associated Steatotic Liver Disease (MASLD) patients. Steatosis is a new silent epidemic and its accurate measurement is an impelling clinical need, not only for hepatologists, but also for experts in metabolic and cardiovascular diseases. This paper aims to provide a robust comparison between different uncertainty quantification strategies to identify advantages and drawbacks in a real clinical setting. Methods We used a classical Convolutional Neural Network, a Monte Carlo Dropout, and a Bayesian Convolutional Neural Network with the goal of not only comparing the goodness of the predictions, but also to have access to an evaluation of the uncertainty associated with the outputs. Results We found that even if the prediction based on a single ultrasound view is reliable (relative RMSE [5.93%-12.04%]), networks based on two ultrasound views outperform them (relative RMSE [5.35%-5.87%]). In addition, the results show that the introduction of a “not confident” category contributes to increase the percentage of correctly predicted cases and to decrease the percentage of mispredicted cases, especially for semi-intrusive methods. Conclusions The possibility of having access to information about the confidence with which the network produces its outputs is a great advantage, both from the point of view of physicians who want to use neural networks as computer-aided diagnosis, and for developers who want to limit overfitting and obtain information about dataset problems in terms of out-of-distribution detection.Source: COMPUTATIONAL AND STRUCTURAL BIOTECHNOLOGY JOURNAL, vol. 24, pp. 603-610
DOI: 10.1016/j.csbj.2024.09.021
Metrics:
G. Del Corso, M. A. Pascali, C. Caudai, L. De Rosa, A. Salvati, M. Mancini, L. Ghiadoni, F. Bonino, M. R. Brunetto, S. Colantonio, F. Faita
Background and objective This article uses three different probabilistic convolutional architectures applied to ultrasound image analysis for grading Fatty Liver Content (FLC) in Metabolic Dysfunction Associated Steatotic Liver Disease (MASLD) patients. Steatosis is a new silent epidemic and its accurate measurement is an impelling clinical need, not only for hepatologists, but also for experts in metabolic and cardiovascular diseases. This paper aims to provide a robust comparison between different uncertainty quantification strategies to identify advantages and drawbacks in a real clinical setting. Methods We used a classical Convolutional Neural Network, a Monte Carlo Dropout, and a Bayesian Convolutional Neural Network with the goal of not only comparing the goodness of the predictions, but also to have access to an evaluation of the uncertainty associated with the outputs. Results We found that even if the prediction based on a single ultrasound view is reliable (relative RMSE [5.93%-12.04%]), networks based on two ultrasound views outperform them (relative RMSE [5.35%-5.87%]). In addition, the results show that the introduction of a “not confident” category contributes to increase the percentage of correctly predicted cases and to decrease the percentage of mispredicted cases, especially for semi-intrusive methods. Conclusions The possibility of having access to information about the confidence with which the network produces its outputs is a great advantage, both from the point of view of physicians who want to use neural networks as computer-aided diagnosis, and for developers who want to limit overfitting and obtain information about dataset problems in terms of out-of-distribution detection.Source: COMPUTATIONAL AND STRUCTURAL BIOTECHNOLOGY JOURNAL, vol. 24, pp. 603-610
DOI: 10.1016/j.csbj.2024.09.021
Metrics:
See at:
Computational and Structural Biotechnology Journal | IRIS - Institutional Research Information System of the University of Trento | IRIS Cnr | IRIS Cnr | IRIS - Institutional Research Information System of the University of Trento | IRIS - Institutional Research Information System of the University of Trento | CNR IRIS
2024
Other
Open Access
U-ProBE: a graphical Python interface to handle uncertainties in deep learning models
Bandini L., Bacciu D., Del Corso G., Caudai C.
La piattaforma U-Probe mira ad essere un supporto ad un crescente numero di utilizzatori che hanno necessità di valutare le performances di un modello e soprat- tutto la sua incertezza. Per questo motivo la piattaforma è dotata di una intuitiva interfaccia grafica semplice da utilizzare anche per i non addetti ai lavori. L’analisi dell’incertezza delle predizioni di un modello di Machine Learning o Deep Learning può essere effettuata utilizzando varie tecniche. Alcune di queste sono intrusive (anche dette by design), tali tecniche vanno a modificare l’architettura in- troducendo strumenti probabilistici che possono fornire importanti indicazioni sulle caratteristiche delle predizioni, a livello di affidabilità e incertezza. Tali tecniche comprendono ad esempio le Bayesian Neural Networks, i Variational Autoencoders ed i Deep Gaussian Processes. Sono tecniche molto performanti, sia nel mitigare l’overfitting che nell’uncertainty quantification, di contro sono però molto costose e richiedono molte risorse di calcolo e di tempo per l’allenamento dei modelli. Esistono poi le tecniche semi-intrusive, i cui più conosciuti rappresentanti sono i Deep En- semble; esse rappresentano una ampia classe di approcci che in generale combinano più modelli secondo criteri specifici in modo da valutare l’efficienza, l’incertezza e l’affidabilità delle predizioni senza interferire troppo con le architetture di partenza, ma richiedendo comunque un ampio dispendio di risorse. In questo lavoro abbiamo deciso di utilizzare per i nostri scopi esclusivamente metodi post-hoc, cioè non intrusivi, come il Trust Score ed il Monte Carlo Dropout, che sono in grado di fare efficaci valutazioni sull’incertezza delle predizioni quando il modello è stato già allenato, senza andare a interferire con le fasi di apprendimento o a modificare i parametri già imparati dal modello durante la back propagation. Tali metodi sono leggermente meno performanti dei metodi intrusivi, ma hanno il vantaggio di essere estremamente più rapidi e meno costosi.
Bandini L., Bacciu D., Del Corso G., Caudai C.
La piattaforma U-Probe mira ad essere un supporto ad un crescente numero di utilizzatori che hanno necessità di valutare le performances di un modello e soprat- tutto la sua incertezza. Per questo motivo la piattaforma è dotata di una intuitiva interfaccia grafica semplice da utilizzare anche per i non addetti ai lavori. L’analisi dell’incertezza delle predizioni di un modello di Machine Learning o Deep Learning può essere effettuata utilizzando varie tecniche. Alcune di queste sono intrusive (anche dette by design), tali tecniche vanno a modificare l’architettura in- troducendo strumenti probabilistici che possono fornire importanti indicazioni sulle caratteristiche delle predizioni, a livello di affidabilità e incertezza. Tali tecniche comprendono ad esempio le Bayesian Neural Networks, i Variational Autoencoders ed i Deep Gaussian Processes. Sono tecniche molto performanti, sia nel mitigare l’overfitting che nell’uncertainty quantification, di contro sono però molto costose e richiedono molte risorse di calcolo e di tempo per l’allenamento dei modelli. Esistono poi le tecniche semi-intrusive, i cui più conosciuti rappresentanti sono i Deep En- semble; esse rappresentano una ampia classe di approcci che in generale combinano più modelli secondo criteri specifici in modo da valutare l’efficienza, l’incertezza e l’affidabilità delle predizioni senza interferire troppo con le architetture di partenza, ma richiedendo comunque un ampio dispendio di risorse. In questo lavoro abbiamo deciso di utilizzare per i nostri scopi esclusivamente metodi post-hoc, cioè non intrusivi, come il Trust Score ed il Monte Carlo Dropout, che sono in grado di fare efficaci valutazioni sull’incertezza delle predizioni quando il modello è stato già allenato, senza andare a interferire con le fasi di apprendimento o a modificare i parametri già imparati dal modello durante la back propagation. Tali metodi sono leggermente meno performanti dei metodi intrusivi, ma hanno il vantaggio di essere estremamente più rapidi e meno costosi.
2024
Journal article
Open Access
Association between Ustekinumab trough levels, Serum IL-22, and Oncostatin M levels and clinical and biochemical outcomes in patients with Crohn’s Disease
Bertin L., Barberio B., Gubbiotti A., Bertani L., Costa F., Ceccarelli L., Visaggi P., Bodini G., Pasta A., Sablich R., Urbano M. T., Ferronato A., Buda A., De Bona M., Del Corso G., Massano G., Angriman I., Scarpa M., Zingone F., Savarino E.
Background: Ustekinumab (UST) has demonstrated effectiveness in treating patients with Crohn's disease. Monitoring treatment response can improve disease management and reduce healthcare costs. We investigated whether UST trough levels (TLs), serum IL22, and Oncostatin M (OSM) levels could be early indicators of non-response by analysing their correlation with clinical and biochemical outcomes in CD. Methods: Patients with CD initiating UST treatment from October 2018 to September 2020 were enrolled at six Italian centres for inflammatory bowel disease (IBD). Clinical and biochemical data were collected at four time points: baseline, second subcutaneous (SC) dose, fourth SC dose, and 52 weeks. TLs were measured during maintenance, at the second SC dose, and at the fourth SC dose. IL-22 and OSM serum levels were assessed at baseline and the second SC dose. We analysed whether TLs, IL22 levels, and OSM serum levels were associated with clinical response, clinical remission, biochemical remission, and endoscopic remission using the appropriate statistical tests. Results: Out of eighty-four initially enrolled patients, five were lost to follow-up, and eleven discontinued the drug before 52 weeks. At the 52-week time point, 47% achieved biochemical remission based on faecal calprotectin levels, and 61.8% achieved clinical remission. TLs at the second SC dose significantly correlated with biochemical remission at the same time point (p = 0.011). However, TLs did not correlate with clinical remission. Baseline OSM levels did not correlate with biochemical or clinical remission or response. IL22 levels notably decreased during UST therapy (p = 0.000), but its values did not correlate with biochemical or clinical remission. Conclusions: UST is an effective therapy for patients with CD. TLs measured at the second SC dose significantly correlated with biochemical remission, emphasising their potential role in treatment monitoring. Levels of OSM and IL-22, despite a significant decrease in the latter during therapy, did not exhibit correlations with clinical or biochemical outcomes in our study. Further studies are needed to confirm these findings.Source: JOURNAL OF CLINICAL MEDICINE, vol. 13 (issue 6)
DOI: 10.3390/jcm13061539
Metrics:
Bertin L., Barberio B., Gubbiotti A., Bertani L., Costa F., Ceccarelli L., Visaggi P., Bodini G., Pasta A., Sablich R., Urbano M. T., Ferronato A., Buda A., De Bona M., Del Corso G., Massano G., Angriman I., Scarpa M., Zingone F., Savarino E.
Background: Ustekinumab (UST) has demonstrated effectiveness in treating patients with Crohn's disease. Monitoring treatment response can improve disease management and reduce healthcare costs. We investigated whether UST trough levels (TLs), serum IL22, and Oncostatin M (OSM) levels could be early indicators of non-response by analysing their correlation with clinical and biochemical outcomes in CD. Methods: Patients with CD initiating UST treatment from October 2018 to September 2020 were enrolled at six Italian centres for inflammatory bowel disease (IBD). Clinical and biochemical data were collected at four time points: baseline, second subcutaneous (SC) dose, fourth SC dose, and 52 weeks. TLs were measured during maintenance, at the second SC dose, and at the fourth SC dose. IL-22 and OSM serum levels were assessed at baseline and the second SC dose. We analysed whether TLs, IL22 levels, and OSM serum levels were associated with clinical response, clinical remission, biochemical remission, and endoscopic remission using the appropriate statistical tests. Results: Out of eighty-four initially enrolled patients, five were lost to follow-up, and eleven discontinued the drug before 52 weeks. At the 52-week time point, 47% achieved biochemical remission based on faecal calprotectin levels, and 61.8% achieved clinical remission. TLs at the second SC dose significantly correlated with biochemical remission at the same time point (p = 0.011). However, TLs did not correlate with clinical remission. Baseline OSM levels did not correlate with biochemical or clinical remission or response. IL22 levels notably decreased during UST therapy (p = 0.000), but its values did not correlate with biochemical or clinical remission. Conclusions: UST is an effective therapy for patients with CD. TLs measured at the second SC dose significantly correlated with biochemical remission, emphasising their potential role in treatment monitoring. Levels of OSM and IL-22, despite a significant decrease in the latter during therapy, did not exhibit correlations with clinical or biochemical outcomes in our study. Further studies are needed to confirm these findings.Source: JOURNAL OF CLINICAL MEDICINE, vol. 13 (issue 6)
DOI: 10.3390/jcm13061539
Metrics:
See at:
Journal of Clinical Medicine | Archivio istituzionale della ricerca - Università di Padova | IRIS Cnr | IRIS Cnr | CNR IRIS
2024
Journal article
Open Access
Adaptive machine learning approach for importance evaluation of multimodal breast cancer radiomic features
Del Corso G., Germanese D., Caudai C., Anastasi G., Belli P., Formica A., Nicolucci A., Palma S., Pascali M. A., Pieroni S., Trombadori C., Colantonio S., Franchini M., Molinaro S.
Breast cancer holds the highest diagnosis rate among female tumors and is the leading cause of death among women. Quantitative analysis of radiological images shows the potential to address several medical challenges, including the early detection and classification of breast tumors. In the P.I.N.K study, 66 women were enrolled. Their paired Automated Breast Volume Scanner (ABVS) and Digital Breast Tomosynthesis (DBT) images, annotated with cancerous lesions, populated the first ABVS+DBT dataset. This enabled not only a radiomic analysis for the malignant vs. benign breast cancer classification, but also the comparison of the two modalities. For this purpose, the models were trained using a leave-one-out nested cross-validation strategy combined with a proper threshold selection approach. This approach provides statistically significant results even with medium-sized data sets. Additionally it provides distributional variables of importance, thus identifying the most informative radiomic features. The analysis proved the predictive capacity of radiomic models even using a reduced number of features. Indeed, from tomography we achieved AUC-ROC 89.9% using 19 features and 92.1% using 7 of them; while from ABVS we attained an AUC-ROC of 72.3% using 22 features and 85.8% using only 3 features. Although the predictive power of DBT outperforms ABVS, when comparing the predictions at the patient level, only 8.7% of lesions are misclassified by both methods, suggesting a partial complementarity. Notably, promising results (AUC-ROC ABVS-DBT 71.8% - 74.1% ) were achieved using non-geometric features, thus opening the way to the integration of virtual biopsy in medical routine.Source: JOURNAL OF IMAGING INFORMATICS IN MEDICINE, vol. 37 (issue 4), pp. 1642-1651
DOI: 10.1007/s10278-024-01064-3
Project(s): "Mortalità Zero - verso la personalizzazione degli interventi diagnostici"
Metrics:
Del Corso G., Germanese D., Caudai C., Anastasi G., Belli P., Formica A., Nicolucci A., Palma S., Pascali M. A., Pieroni S., Trombadori C., Colantonio S., Franchini M., Molinaro S.
Breast cancer holds the highest diagnosis rate among female tumors and is the leading cause of death among women. Quantitative analysis of radiological images shows the potential to address several medical challenges, including the early detection and classification of breast tumors. In the P.I.N.K study, 66 women were enrolled. Their paired Automated Breast Volume Scanner (ABVS) and Digital Breast Tomosynthesis (DBT) images, annotated with cancerous lesions, populated the first ABVS+DBT dataset. This enabled not only a radiomic analysis for the malignant vs. benign breast cancer classification, but also the comparison of the two modalities. For this purpose, the models were trained using a leave-one-out nested cross-validation strategy combined with a proper threshold selection approach. This approach provides statistically significant results even with medium-sized data sets. Additionally it provides distributional variables of importance, thus identifying the most informative radiomic features. The analysis proved the predictive capacity of radiomic models even using a reduced number of features. Indeed, from tomography we achieved AUC-ROC 89.9% using 19 features and 92.1% using 7 of them; while from ABVS we attained an AUC-ROC of 72.3% using 22 features and 85.8% using only 3 features. Although the predictive power of DBT outperforms ABVS, when comparing the predictions at the patient level, only 8.7% of lesions are misclassified by both methods, suggesting a partial complementarity. Notably, promising results (AUC-ROC ABVS-DBT 71.8% - 74.1% ) were achieved using non-geometric features, thus opening the way to the integration of virtual biopsy in medical routine.Source: JOURNAL OF IMAGING INFORMATICS IN MEDICINE, vol. 37 (issue 4), pp. 1642-1651
DOI: 10.1007/s10278-024-01064-3
Project(s): "Mortalità Zero - verso la personalizzazione degli interventi diagnostici"
Metrics:
See at:
Journal of Imaging Informatics in Medicine | IRIS Cnr | IRIS Cnr | CNR IRIS | CNR IRIS
2024
Journal article
Open Access
Optimizing radiomics for prostate cancer diagnosis: feature selection strategies, machine learning classifiers, and MRI sequences
Mylona E., Zaridis D. I., Kalantzopoulos C., Tachos N. S., Regge D., Papanikolaou N., Tsiknakis M., Marias K., Marquez R., Henne T., Saillant C., Mora J. M., Pastor A. J., Agraniotis D., Pollalis C., Giavri Z., Hernandez W., Correia J., Bridge C., Kalpathy-Cramer J., Carloni G., Berti A., Germanese D., Del Corso G., Pachetti E., Pascali M. A., Colantonio S., Napolitano V., Maimone G., Cappello G., Mazzetti S., Giannini V., García-Martí G., Jacobs T., Doran S., Ribeiro A., Vit S., Emsley R., Koh D. M., Georgios G., Vasilis K., Slidevska K., Untanas A., Briediene R., Usinskiene J., Vilanova J. C., Karcaaltincaba M., Atak F., Karaosmanoglu A. D., Özmen M., Akata D., Nan, Mendola V., Tumminello L., Aringhieri G., Neri E., Marfil M., Navarro S., Ribas G., Cerdá-Alberich L., Martí-Bonmatí L., Futterer J., Twilt J. J., Saha A., De Rooij M., Huisman H., Chambel M., Rodrigues N., Rodrigues A. C., Verde A. C., De Almeida J. G., Dimitriadis A., Kalliatakis G., Trivizakis E., Kalokyri V., Sfakianakis S., Fotiadis D. I.
Objectives: Radiomics-based analyses encompass multiple steps, leading to ambiguity regarding the optimal approaches for enhancing model performance. This study compares the effect of several feature selection methods, machine learning (ML) classifiers, and sources of radiomic features, on models' performance for the diagnosis of clinically significant prostate cancer (csPCa) from bi-parametric MRI. Methods: Two multi-centric datasets, with 465 and 204 patients each, were used to extract 1246 radiomic features per patient and MRI sequence. Ten feature selection methods, such as Boruta, mRMRe, ReliefF, recursive feature elimination (RFE), random forest (RF) variable importance, L1-lasso, etc., four ML classifiers, namely SVM, RF, LASSO, and boosted generalized linear model (GLM), and three sets of radiomics features, derived from T2w images, ADC maps, and their combination, were used to develop predictive models of csPCa. Their performance was evaluated in a nested cross-validation and externally, using seven performance metrics. Results: In total, 480 models were developed. In nested cross-validation, the best model combined Boruta with Boosted GLM (AUC = 0.71, F1 = 0.76). In external validation, the best model combined L1-lasso with boosted GLM (AUC = 0.71, F1 = 0.47). Overall, Boruta, RFE, L1-lasso, and RF variable importance were the top-performing feature selection methods, while the choice of ML classifier didn't significantly affect the results. The ADC-derived features showed the highest discriminatory power with T2w-derived features being less informative, while their combination did not lead to improved performance. Conclusion: The choice of feature selection method and the source of radiomic features have a profound effect on the models' performance for csPCa diagnosis. Critical relevance statement: This work may guide future radiomic research, paving the way for the development of more effective and reliable radiomic models; not only for advancing prostate cancer diagnostic strategies, but also for informing broader applications of radiomics in different medical contexts. Key points: Radiomics is a growing field that can still be optimized. Feature selection method impacts radiomics models' performance more than ML algorithms. Best feature selection methods: RFE, LASSO, RF, and Boruta. ADC-derived radiomic features yield more robust models compared to T2w-derived radiomic features.Source: INSIGHTS INTO IMAGING, vol. 15 (issue 1)
DOI: 10.1186/s13244-024-01783-9
Project(s): ProCAncer-I
Metrics:
Mylona E., Zaridis D. I., Kalantzopoulos C., Tachos N. S., Regge D., Papanikolaou N., Tsiknakis M., Marias K., Marquez R., Henne T., Saillant C., Mora J. M., Pastor A. J., Agraniotis D., Pollalis C., Giavri Z., Hernandez W., Correia J., Bridge C., Kalpathy-Cramer J., Carloni G., Berti A., Germanese D., Del Corso G., Pachetti E., Pascali M. A., Colantonio S., Napolitano V., Maimone G., Cappello G., Mazzetti S., Giannini V., García-Martí G., Jacobs T., Doran S., Ribeiro A., Vit S., Emsley R., Koh D. M., Georgios G., Vasilis K., Slidevska K., Untanas A., Briediene R., Usinskiene J., Vilanova J. C., Karcaaltincaba M., Atak F., Karaosmanoglu A. D., Özmen M., Akata D., Nan, Mendola V., Tumminello L., Aringhieri G., Neri E., Marfil M., Navarro S., Ribas G., Cerdá-Alberich L., Martí-Bonmatí L., Futterer J., Twilt J. J., Saha A., De Rooij M., Huisman H., Chambel M., Rodrigues N., Rodrigues A. C., Verde A. C., De Almeida J. G., Dimitriadis A., Kalliatakis G., Trivizakis E., Kalokyri V., Sfakianakis S., Fotiadis D. I.
Objectives: Radiomics-based analyses encompass multiple steps, leading to ambiguity regarding the optimal approaches for enhancing model performance. This study compares the effect of several feature selection methods, machine learning (ML) classifiers, and sources of radiomic features, on models' performance for the diagnosis of clinically significant prostate cancer (csPCa) from bi-parametric MRI. Methods: Two multi-centric datasets, with 465 and 204 patients each, were used to extract 1246 radiomic features per patient and MRI sequence. Ten feature selection methods, such as Boruta, mRMRe, ReliefF, recursive feature elimination (RFE), random forest (RF) variable importance, L1-lasso, etc., four ML classifiers, namely SVM, RF, LASSO, and boosted generalized linear model (GLM), and three sets of radiomics features, derived from T2w images, ADC maps, and their combination, were used to develop predictive models of csPCa. Their performance was evaluated in a nested cross-validation and externally, using seven performance metrics. Results: In total, 480 models were developed. In nested cross-validation, the best model combined Boruta with Boosted GLM (AUC = 0.71, F1 = 0.76). In external validation, the best model combined L1-lasso with boosted GLM (AUC = 0.71, F1 = 0.47). Overall, Boruta, RFE, L1-lasso, and RF variable importance were the top-performing feature selection methods, while the choice of ML classifier didn't significantly affect the results. The ADC-derived features showed the highest discriminatory power with T2w-derived features being less informative, while their combination did not lead to improved performance. Conclusion: The choice of feature selection method and the source of radiomic features have a profound effect on the models' performance for csPCa diagnosis. Critical relevance statement: This work may guide future radiomic research, paving the way for the development of more effective and reliable radiomic models; not only for advancing prostate cancer diagnostic strategies, but also for informing broader applications of radiomics in different medical contexts. Key points: Radiomics is a growing field that can still be optimized. Feature selection method impacts radiomics models' performance more than ML algorithms. Best feature selection methods: RFE, LASSO, RF, and Boruta. ADC-derived radiomic features yield more robust models compared to T2w-derived radiomic features.Source: INSIGHTS INTO IMAGING, vol. 15 (issue 1)
DOI: 10.1186/s13244-024-01783-9
Project(s): ProCAncer-I
Metrics:
See at:
Insights into Imaging | IRIS Cnr | CNR IRIS | insightsimaging.springeropen.com | IRIS Cnr | IRIS Cnr | IRIS Cnr | CNR IRIS
2024
Other
Embargo
Monitoraggio della termoregolazione neonatale in contesto ospedaliero: verso un approccio integrato e non invasivo
Cancello Tortora C., Del Corso G., Germanese D., Positano V., Vozzi G.
La termoregolazione, ovvero la capacità di mantenere una temperatura adeguata, è una questione di notevole interesse e complessità nella comunità scientifica nell'ambito della neonatologia. Nei primi istanti di vita i neonati, sia pre-termine che a termine, presentano sistemi di regolazione della temperatura immaturi che li rendono vulnerabili alle condizioni subottimali extra-uterine. Questa tesi, svolta presso il Laboratorio Segnali e Immagini dell'Istituto di Scienze e Tecnologie dell'Informazione del CNR di Pisa, propone di monitorare, in contesto ospedaliero, mediante un sistema integrato e non invasivo, le variazioni di temperatura del neonato nelle prime ore di vita. L’obiettivo è quello di preparare il terreno per uno studio più ampio che, attraverso l’acquisizione e la valutazione di pattern termici sul neonato, sarà in grado di valutare lo stato patologico del neonato. Inoltre, verrà valutato se la stabilizzazione termica possa essere migliorata con l’attuazione di una pratica, nota come contatto pelle a pelle (SSC), tra madre e neonato, o eventualmente tra padre e neonato. L’ hardware del dispositivo è stato realizzato dal Centro di Formazione e Simulazione Neonatale (centro NINA) dell'Azienda Ospedaliero Universitaria Pisana. L'idea di monitorare i pattern termici di un neonato in maniera non invasiva si è tradotta in un dispositivo estremamente compatto e portatile, costituito da: (i) una termocamera, mediante la quale acquisire le immagini termiche del neonato, (ii) una telecamera rgb per acquisire le immagini del neonato nello spettro del visibile, estrarre lo scheletro per definire automaticamente i distretti anatomici di interesse, (iii) un sensore per la misurazione puntuale della temperatura, (iv) un sensore di umidità e temperatura ambientale per monitorare le condizioni ambientali della stanza in cui si trova il neonato, (v) un Raspeberry Pi per la gestione e l'integrazione di questi componenti nonchè l'estrazione e la pre-elaborazione dei dati. Il Software di controllo ed elaborazione sviluppato in questa tesi è stato scritto in linguaggio Python (v. 3.11) e gestisce gli stati del sistema, in particolare l’acquisizione sincrona di immagini termiche ed RGB, l’estrazione di dati e l’anonimizzazione delle immagini RGB dei neonati. L’elaborazione delle immagini RGB viene effettuata in locale dal Raspberry e comprende l’estrazione automatica delle regioni anatomiche di interesse (ROI) mediante tecniche allo stato dell’arte (i.e., libreria MediaPipe). Successivamente, queste ROI vengono trasposte sulle corrispondenti immagini termiche tramite una matrice di trasformazione omografica opportunamente calibrata tenendo in considerazione il vincolo rigido tra le due camere e le rispettive distanze focali. Queste ROI prendono come riferimento per il punto centrale il landmark estratto e come raggio le proporzioni tra due landmarks vicini e le dimensioni stimate del distretto anatomico di interesse. Esse rappresentano il punto di partenza dell’elaborazione delle immagini termiche. Dopo una fase iniziale di pre-processing, in cui il rumore di fondo è stato eliminato con varie tecniche di filtraggio, il contrasto tra le varie regioni è stato aumentato. Questo processo è stato propedeutico all’estrazione degli istogrammi, il cui andamento fornisce informazioni sulla presenza o meno di sfondo. Se lo sfondo è presente, viene avviato il segmentatore FastSAM, basato su una rete neurale convolutiva (CNN) allo stato dell'arte, che segmenta il distretto anatomico per evitare di includere lo sfondo nell’elaborazione. Un’interfaccia utente user-friendly ha permesso di gestire i landmarks provenienti dallo scheletro e di realizzare in maniera completamente automatica delle regioni di interesse (ROI) adattive sull’immagine termica. Dalla singola ROI sono state estratte dei pattern termici e delle features che estendessero quelle tradizionali come mediana e intervallo interquartile attraverso l’implementazione di una matrice di texture che deriva da descrittori matematici quantitativi di texture (della famiglia GLSZM- Gray level size zone matrix) che forniscono informazioni sull’eterogeneità termica delle ROI. La matrice è stata utile per estrarre un punteggio (score) da attribuire alla singola ROI evidenziando come un paziente con vaste aree di temperatura accettabile avesse un punteggio maggiore rispetto ad un paziente con zone molte fredde ed un’alta variabilità nella temperatura. Infine, sono state definite anche delle features a livello globale che mettono in relazione le misure ottenute dalla ROI sul viso (riferimento clinico neonatale) con quelle sul torace e sugli arti. Il sistema è stato validato prima in un contesto sperimentale controllato, la validazione finale e la conseguente acquisizione di dati sono avvenute in ambito ospedaliero, nel reparto di neonatologia dell'Azienda Ospedaliera Universitaria Pisana, utilizzando un fantoccio che simulava il comportamento termico di un neonato.
Cancello Tortora C., Del Corso G., Germanese D., Positano V., Vozzi G.
La termoregolazione, ovvero la capacità di mantenere una temperatura adeguata, è una questione di notevole interesse e complessità nella comunità scientifica nell'ambito della neonatologia. Nei primi istanti di vita i neonati, sia pre-termine che a termine, presentano sistemi di regolazione della temperatura immaturi che li rendono vulnerabili alle condizioni subottimali extra-uterine. Questa tesi, svolta presso il Laboratorio Segnali e Immagini dell'Istituto di Scienze e Tecnologie dell'Informazione del CNR di Pisa, propone di monitorare, in contesto ospedaliero, mediante un sistema integrato e non invasivo, le variazioni di temperatura del neonato nelle prime ore di vita. L’obiettivo è quello di preparare il terreno per uno studio più ampio che, attraverso l’acquisizione e la valutazione di pattern termici sul neonato, sarà in grado di valutare lo stato patologico del neonato. Inoltre, verrà valutato se la stabilizzazione termica possa essere migliorata con l’attuazione di una pratica, nota come contatto pelle a pelle (SSC), tra madre e neonato, o eventualmente tra padre e neonato. L’ hardware del dispositivo è stato realizzato dal Centro di Formazione e Simulazione Neonatale (centro NINA) dell'Azienda Ospedaliero Universitaria Pisana. L'idea di monitorare i pattern termici di un neonato in maniera non invasiva si è tradotta in un dispositivo estremamente compatto e portatile, costituito da: (i) una termocamera, mediante la quale acquisire le immagini termiche del neonato, (ii) una telecamera rgb per acquisire le immagini del neonato nello spettro del visibile, estrarre lo scheletro per definire automaticamente i distretti anatomici di interesse, (iii) un sensore per la misurazione puntuale della temperatura, (iv) un sensore di umidità e temperatura ambientale per monitorare le condizioni ambientali della stanza in cui si trova il neonato, (v) un Raspeberry Pi per la gestione e l'integrazione di questi componenti nonchè l'estrazione e la pre-elaborazione dei dati. Il Software di controllo ed elaborazione sviluppato in questa tesi è stato scritto in linguaggio Python (v. 3.11) e gestisce gli stati del sistema, in particolare l’acquisizione sincrona di immagini termiche ed RGB, l’estrazione di dati e l’anonimizzazione delle immagini RGB dei neonati. L’elaborazione delle immagini RGB viene effettuata in locale dal Raspberry e comprende l’estrazione automatica delle regioni anatomiche di interesse (ROI) mediante tecniche allo stato dell’arte (i.e., libreria MediaPipe). Successivamente, queste ROI vengono trasposte sulle corrispondenti immagini termiche tramite una matrice di trasformazione omografica opportunamente calibrata tenendo in considerazione il vincolo rigido tra le due camere e le rispettive distanze focali. Queste ROI prendono come riferimento per il punto centrale il landmark estratto e come raggio le proporzioni tra due landmarks vicini e le dimensioni stimate del distretto anatomico di interesse. Esse rappresentano il punto di partenza dell’elaborazione delle immagini termiche. Dopo una fase iniziale di pre-processing, in cui il rumore di fondo è stato eliminato con varie tecniche di filtraggio, il contrasto tra le varie regioni è stato aumentato. Questo processo è stato propedeutico all’estrazione degli istogrammi, il cui andamento fornisce informazioni sulla presenza o meno di sfondo. Se lo sfondo è presente, viene avviato il segmentatore FastSAM, basato su una rete neurale convolutiva (CNN) allo stato dell'arte, che segmenta il distretto anatomico per evitare di includere lo sfondo nell’elaborazione. Un’interfaccia utente user-friendly ha permesso di gestire i landmarks provenienti dallo scheletro e di realizzare in maniera completamente automatica delle regioni di interesse (ROI) adattive sull’immagine termica. Dalla singola ROI sono state estratte dei pattern termici e delle features che estendessero quelle tradizionali come mediana e intervallo interquartile attraverso l’implementazione di una matrice di texture che deriva da descrittori matematici quantitativi di texture (della famiglia GLSZM- Gray level size zone matrix) che forniscono informazioni sull’eterogeneità termica delle ROI. La matrice è stata utile per estrarre un punteggio (score) da attribuire alla singola ROI evidenziando come un paziente con vaste aree di temperatura accettabile avesse un punteggio maggiore rispetto ad un paziente con zone molte fredde ed un’alta variabilità nella temperatura. Infine, sono state definite anche delle features a livello globale che mettono in relazione le misure ottenute dalla ROI sul viso (riferimento clinico neonatale) con quelle sul torace e sugli arti. Il sistema è stato validato prima in un contesto sperimentale controllato, la validazione finale e la conseguente acquisizione di dati sono avvenute in ambito ospedaliero, nel reparto di neonatologia dell'Azienda Ospedaliera Universitaria Pisana, utilizzando un fantoccio che simulava il comportamento termico di un neonato.