Analisi di immagini tomografiche ad alta risoluzione attraverso reti neurali convoluzionali per lo studio delle interstiziopatie polmonari

Buongiorno R.

reti neurali convoluzionali  deep learning  interstiziopatie polmonari  elaborazione di immagini biomediche  tomografia computerizzata ad alta risoluzione 

The term Interstitial Lung Disease (ILD) refers to a large group of lung disorders, most of which cause scars of the interstitium, usually referred to as pulmonary fibrosis. Fibrosis reduces the ability of the air sacs to capture and carry oxygen into the bloodstream, leading to a progressive loss of the ability to breathe. Although ILDs are rare if taken individually, together they represent the most frequent cause of non-obstructive chronic lung disease. Nowadays, there are more than 200 different types of ILDs with varying causes, prognosis and therapies. Thus, identifying the correct type of ILD is necessary to make an accurate diagnosis. The Idiopathic Pulmonary Fibrosis (IPF) is a chronic, progressive fibrosing interstitial pneumonia, which is classified among the ILDs with the poorest prognosis. The high variability and unpredictability of IPF course have traditionally made its clinical management hard. The recent introduction of antifibrotic drugs has opened novel therapeutic options for mild to moderate IPF. In this respect, treatment decisions highly rely on the assessment and quantification of IPF impact on the interstitium and its progression over time. High-Resolution Computed Tomography (HRCT) has demonstrated to have a key role in this frame, as it represents a non-invasive diagnostic modality to evaluate and quantify the extent of lung interstitium affected by IPF. In fact, IPF shows a typical radiological pattern, called Usual Interstitial Pneumonia (UIP) pattern, whose presence is usually assessed by radiologists to diagnose IPF. The HRCT features that characterize the UIP pattern are the presence and positioning of specific lung parenchymal anomalies, known as honeycombing , ground-glass opacification and fine reticulation. These anomalies appear in the HRCT scans with specific textural characteristics that are detected via a visual inspection of the imaging data. Assessing the diffusion of these anomalies is instrumental to understand the impact of IPF and to monitor its evolution over time. Quantitative and reliable approaches are in high demand in this respect, as the visual examination by radiologists suffers, by its nature, of poor reproducibility. To overcome this issue, much research is being conducted to develop new techniques for automatic detection of lung diseases that may support radiologists during the diagnostic pathway, particularly in HRCT image analysis. Indeed, HRCT images evaluation by a Machine Learning (ML)- based algorithm might provide low-cost, reliable, real time automatic identification of UIP pattern with human-level accuracy in order to objectively quantify the percentage of lung volume affected by the disease in a reproducible way. The purpose of this study was to develop a tool for UIP pattern recognition in HRCT images of patients with IPF using a deep-learning method based on a Convolutional Neural Network (CNN), called UIP-net. UIP-net takes as input a lung HRCT image with 492x492 pixels and outputs the corresponding binary map for the discrimination of disease and normal tissue. To train and evaluate the CNN, a dataset of 5000 images, derived by 20 CT scans from the same scanner, was used. The network performance yielded 83.7% BF-score and 84.6% sensitivity but in order to refine the binary masks produced by UIP-net, a post-processing operation was carried out. With post-processing, vessels, air-ways and tissue wrongly classified as belonging to the lungs were removed from the outputted masks. After post-processing, the results increased to 96.7% BF-score and 85.9% sensitivity. Thus, the network performance, in terms of BF-score and sensitivity, demonstrated that CNNs have the potential to reliably detect disease in order to evaluate its progression and become a supportive tool for radiologists. Future works include adding more data to the training set in order to add multiple layers to the network to distinguish and quantify the different HRCT features of UIP pattern, improving the reproducibility and reliability of the CNN and using it for the detection of HRCT manifestations of Covid-19.