Automatic CT Lesion Detection Based on Feature Pyramid Inference with Multi-scale Response.

In modern medical applications, computed tomography image is used as one of the most useful tools for diagnosis and localization of lesions. It can provide patients with precise information about the location and size of their tumor lesions. Traditional medical diagnosis is not only very time consuming but also not very accurate. Nowadays, the automatic detection of lesions on computed tomography has become a research area of great interest, and researchers aim to use computer-aided diagnosis to assist in clinical medical diagnosis. However, for current detection algorithms, the accuracy of automatic lesion detection is still low, especially for small lesions. In this paper, to improve the accuracy of detection of small lesions, we propose a Multi-Scale Response Module (MSR) that incorporates global attention into Feature Pyramid Network (FPN) build on backbone. At each pyramid level, the proposed Aggregated Dilation Block (ADB) is used to capture the variations in the fine-grained scales. The response of the network to small lesion features is then reinforced by the Global Attention Block (GAB). We build a Feature Pyramid Network (FPN) based on the highly responsive output of the MSR module, with each layer of the FPN fusing high semantic information from low resolution layers. The experimental results show that our method has a higher detection accuracy with mAP value of 58.4 and a high sensitivity compared to the state-of-the-art methods.