Diagnosis of Occupational Pneumoconiosis Based on Adaptive Feature Fusion and Multi-Class Loss Function

Pneumoconiosis, as the most prevalent and severe occupational disease in China, is categorized into four stages based on the area of small shadows in lungs. Previous diagnostic methods have limited accuracy for identifying different stages due to complex and diverse lesions of pneumoconiosis, as well as the difficulty of collecting rich X-ray images of patients. To address the issues, a novel diagnostic method based on adaptive feature fusion and multi-class loss function is proposed. First, the prototype network is introduced to establish a diagnostic model under the insufficient images, with the purpose of distinguishing the disease stage by the distance between the features of various samples. Following that, the multi-channel and multi-level feature fusion modules are proposed, respectively. These modules incorporate a gate-based adaptive selection mechanism to efficiently integrate the most important features from different channels and levels. Finally, the multi-class loss function is designed to train the diagnostic model of pneumoconiosis. The experimental results on a self-built dataset of chest X-ray radiographs demonstrate that the proposed method outperforms ten comparative methods in terms of comprehensive performance.