Loose Lesion Location Self-supervision Enhanced Colorectal Cancer Diagnosis

Tianhong Gao, Jie Song, Xiaotian Yu, Shengxuming Zhang, Wenjie Liang, Hongbin Zhang, Ziqian Li, Wenzhuo Zhang, Xiuming Zhang, Zipeng Zhong, Mingli Song, Zunlei Feng

Abstract: Early diagnosis of colorectal cancer (CRC) is crucial for improving survival and quality of life. While computed tomography (CT) is a key diagnostic tool, manually screening colon tumors is time-consuming and repetitive for radiologists. Recently, deep learning has shown promise in medical image analysis, but its clinical application is limited by the model’s unexplainability and the need for a large number of finely annotated samples. In this paper, we propose a loose lesion location self-supervision enhanced CRC diagnosis framework to reduce the requirement of fine sample annotations and improve the reliability of prediction results. For both non-contrast and contrast CT, despite potential deviations in imaging positions, the lesion location should be nearly consistent in images of both modalities at the same sequence position. In addition, lesion location in two successive slices is relatively close for the same modality. Therefore, a self-supervision mechanism is devised to enforce lesion location consistency at both temporal and modality levels of CT, reducing the need for fine annotations and enhancing the interpretability of diagnostics. Furthermore, this paper introduces a mask correction loopback strategy to reinforce the interdependence between category label and lesion location, ensuring the reliability of diagnosis. To verify our method’s effectiveness, we collect data from 3,178 CRC patients and 887 healthy controls. Experiment results show that the proposed method not only provides reliable lesion localization but also enhances the classification performance by 1–2%, offering an effective diagnostic tool for CRC. Code is available at https://github.com/Gaotianhong/LooseLocationSS.