Computed tomography of interstitial lung disease in systemic sclerosis: dataset and deep learning model for pulmonary lesion segmentation.
Systemic sclerosis-associated interstitial lung disease (SSc-ILD) is a major cause of morbidity and mortality in systemic sclerosis patients. High-resolution computed tomography (HRCT) plays a crucial role in SSc-ILD diagnosis and management. In this study, we aimed to develop and evaluate a deep learning-based automated segmentation model for quantifying SSc-ILD lesions in HRCT images and assess its clinical relevance.
We developed a convolutional neural network model to segment normal lung, established fibrosis (EF), and ground-glass opacity (GGO) in HRCT scans from 40 SSc-ILD patients. The model was trained and evaluated using 8-fold cross-validation. Segmentation performance was assessed using the Dice score coefficient (DSC). Correlations between predicted lesion volumes and pulmonary function test (PFT) metrics were analyzed using Spearman's ρ.
Our model achieved a total lesion (EF+GGO) DSC of 78%. Class-wise segmentation performance was lower for EF (DSC: 70%) compared to GGO (DSC: 73%). Predicted lesion volumes showed significant negative correlations with forced expiratory volume in one second (FEV1) (ρ=-0.64, p<0.001) and FEV1/forced vital capacity (ρ=-0.73, p<0.001). We also created the SICCS dataset, a public dataset of SSc-ILD HRCT images with expert-annotated segmentation labels.
Deep learning-based automated segmentation can help quantify SSc-ILD lesions in HRCT images and provide clinically relevant information. The model's performance is comparable to previous studies, and the predicted lesion volumes correlate significantly with PFT metrics. This approach shows promise for aiding in SSc-ILD diagnosis, monitoring, and clinical decision-making, although further validation with larger datasets is needed.
We developed a convolutional neural network model to segment normal lung, established fibrosis (EF), and ground-glass opacity (GGO) in HRCT scans from 40 SSc-ILD patients. The model was trained and evaluated using 8-fold cross-validation. Segmentation performance was assessed using the Dice score coefficient (DSC). Correlations between predicted lesion volumes and pulmonary function test (PFT) metrics were analyzed using Spearman's ρ.
Our model achieved a total lesion (EF+GGO) DSC of 78%. Class-wise segmentation performance was lower for EF (DSC: 70%) compared to GGO (DSC: 73%). Predicted lesion volumes showed significant negative correlations with forced expiratory volume in one second (FEV1) (ρ=-0.64, p<0.001) and FEV1/forced vital capacity (ρ=-0.73, p<0.001). We also created the SICCS dataset, a public dataset of SSc-ILD HRCT images with expert-annotated segmentation labels.
Deep learning-based automated segmentation can help quantify SSc-ILD lesions in HRCT images and provide clinically relevant information. The model's performance is comparable to previous studies, and the predicted lesion volumes correlate significantly with PFT metrics. This approach shows promise for aiding in SSc-ILD diagnosis, monitoring, and clinical decision-making, although further validation with larger datasets is needed.
Authors
Vahdani Vahdani, Mohamadi Mohamadi, Nayebirad Nayebirad, Forghani Forghani, Karimi Karimi, Vojgani Vojgani, Tamartash Tamartash, Asadi Asadi, Aghaghazvini Aghaghazvini, Gharibdoost Gharibdoost, Kavosi Kavosi
View on Pubmed