Abstract

Info

Title:: Deep learning network for lung cancer detection and localization in PET-CT images under long-tailed distribution

Author(s):: LIANG Zhixin1; DENG Gelong2; WEI Xiaping3; LIANG Fenghao1; HUI Xianjuan1; LIANG Yijun1; LI Qian1; HUANG Xiaowei4; LUO Rongcheng5; 1. Department of Nuclear Medicine, Jinshazhou Hospital, Guangzhou University of Chinese Medicine, Guangzhou 510168, China 2. School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China 3. Department of Radiation Oncology, Jinshazhou Hospital, Guangzhou University of Chinese Medicine, Guangzhou 510168, China 4. Science and Technology Division, Dongguan University of Technology, Dongguan 523808, China 5. Department of Oncology, Jinshazhou Hospital, Guangzhou University of Chinese Medicine, Guangzhou 510168, China

Keywords:: Keywords: lung cancer PET/CT long-tailed distribution deep learning target detection

Abstract:: Abstract: Objective To propose a method for lung cancer detection in PET/CT images under long-tailed distribution for improving the efficiency of PET/CT images for lung cancer diagnosis. Methods A YOLO framework called adaptive class loss function-YOLO (ACS-YOLO) which mainly consists of a YOLOv5 as backbone network (Backbone) and an adaptive class loss function (ACSLoss) was constructed to solve the problem of long-tailed distribution in the real data set of PET/CT lung cancer images and to improve the efficiency of PET/CT images for lung cancer diagnosis. Results Among the existing YOLOv5 variants on the Lung-PET/CT-Dx public data set, the proposed ACS-YOLO achieved the best detection performances in the precision, recall, mAP@0.5 and mAP@0.5:0.95 which were 0.960 7, 0.948 9, 0.970 6 and 0.558 3, respectively. Compared with the other YOLOv5 variants, ACS-YOLO showed a 1%-5% improvement in detection performance and a 5%-11% improvement in tail category detection performance. Conclusion The proposed ACS-YOLO can effectively improve the lung cancer detection in PET/CT images under long-tailed distribution, which indicates that the proposed method can be used as an aid for lung cancer diagnosis using PET/CT images.