Deep-learning-based segmentation of interstitial lung disease lesions in chest HRCT

Interstitial Lung Disease (ILD) encompasses over 200 disease subtypes and is a significant category within chronic respiratory diseases. The diagnosis of ILD faces challenges due to its complexity and the risks associated with invasive examinations, heavily relying on High-Resolution Computed Tomography (HRCT). Manual interpretation of these images is limited by the doctor's experience and subjective factors, leading to potential misdiagnoses and requiring extensive time for report generation. Therefore, establishing an effective lesion segmentation method to advance standardized diagnosis and treatment of interstitial lung diseases is necessary. This paper employs a two-stage transfer learning approach to train a Deeplabv3+ model with an atrous convolution spatial pyramid pooling module for segmenting four types of ILD lesions. This method initially learns general features such as edges and textures from natural images, then familiarizes the model with medical image styles, CT densities, and lesion shapes using a low-quality dataset, before fine-tuning on a high-quality, task-consistent dataset to produce the final lung lesion segmentation model.