Lesion attentive thoracic disease diagnosis with large decision margin loss

The automatic analysis of chest X-rays has been widely explored to develop a thoracic disease diagnosis system. Two obstacles affect the performance of conventional deep models. First, the position and size of lung lobes in different X-ray images vary significantly, making lesions different in appearance. Second, X-ray images sometimes contain irrelevant objects such as supporting devices, which contain much interference information, making it difficult to diagnose thoracic diseases accurately. To this end, an improved conventional deep model is aimed in this paper by proposing two novel components, the lesion attentive network, and the large decision margin loss. The proposed lesion attentive network imitates how a radiologist would analyze an image, focusing only on the most disease-related regions of the input images, which improves the efficiency of the model for feature extraction. The proposed large decision margin loss, in turn, encourages the model to excavate more accurate supporting evidence for predictions, increasing its generalization ability. Extensive experiments are conducted on the NIH Chest X-ray dataset. The proposed approach achieves the state-of-the-art thorax disease diagnosis performances on the validation and test set, demonstrating its superiority.

Automated summary

The study proposes an improved deep model with two novel components, the lesion attentive network and the large decision margin loss, aiming to address the obstacles in chest X-ray image analysis. The proposed approach achieves state-of-the-art thorax disease diagnosis performances on the validation and test set, demonstrating its superiority in accuracy and efficiency.