Enhanced Lung Lesion Detection with Improved Attention Mechanisms

Abstract

In recent years, artificial intelligence (AI) technology has been extensively utilized to aid radiologists in the diagnosis and analysis of medical images. AI proves particularly effective in assisting with the localization of lesions. Nevertheless, the current mainstream target detection models face practical implementation challenges within medical systems due to their reliance on large backbone networks and high input resolutions, which result in reduced accuracy and significant computational resource demands. This paper introduces the IAB-RefineDet, a lung lesion detection network characterized by rapid detection speeds and high accuracy. By enhancing the channel and spatial attention mechanisms and integrating an improved attention module into RefineDet, the accuracy of lesion detection is markedly increased without a substantial rise in parameter numbers. We conduct comprehensive experiments on VinDr-CXR, the largest publicly available chest radiograph detection dataset, alongside comparative studies with existing mainstream target detection models. The experimental outcomes demonstrate that IAB-RefineDet achieves a mean Average Precision (mAP) of 16.24%, significantly outperforming leading deep learning models in lesion detection performance.