Automatic assessment of lung involvement in systemic sclerosis using deep learning

Systemic sclerosis (SSc) is a relatively uncommon connective tissue disorder, commonly manifesting as interstitial lung disease (ILD) and affecting both the lung parenchyma and the modification of the space between endothelium and epithelium. Imaging modalities like computed tomography (CT) scans are essential for diagnosing and revealing specific abnormal findings (ILD patterns) in SSc, such as reticulation and Ground-glass opacity (GGO). To enhance diagnostic precision and minimize human error, we leverage deep learning (DL) techniques.

In our study, we collected and annotated a new public dataset from 22 individuals, encompassing 2190 lung CT scan slices. After preprocessing and exclusion of slices without abnormalities, 1777 slices from 17 patients were used for model training and validation, and 413 slices from five patients were reserved for independent testing. We use a specialized U-net model to segment these patterns, categorizing them into reticulation or GGO, and employ an automated algorithm to outline lung areas in each CT slice. The model's objective is to quantify the patient's lung involvement in SSc by calculating the total identified GGO and reticulation areas across all slices and normalizing this by the total lung surface area.

The U-net model shows promising results in segmenting both reticulation and a combination of GGO and reticulation, as indicated by Dice coefficients of 87.22% and 86.20%, respectively. Furthermore, the automated algorithm effectively outlines the lung region in each slice, enabling accurate measurement of lung involvement in SSc patients.

In conclusion, using DL using the U-Net model and an automated algorithm has shown promising results in accurately segmenting and quantifying lung involvement in Scleroderma patients using CT scans.