Document resource
Background Colon cancer is a leading cause of cancer death worldwide, and CT-based tumor segmentation is critical for staging and treatment planning. However, current methods are hindered by low tumor contrast, complex anatomy ( IDDF2026-ABS-0164 Figure 1. Overview of ColSegNet A clinically interactive framework for multi-center colon cancer CTs), and limited annotated data, with DSC below 0.4257—far from clinical needs. We aim to develop a clinically oriented interactive segmentation model that achieves high accuracy with minimal clinician effort, enabling seamless integration into routine workflows and supporting downstream tasks such as N-stage and MMR prediction.Methods We constructed a large multicenter dataset of 1,290 meticulously annotated colon cancer CT cases from eight independent sources, covering diverse imaging protocols, tumor morphologies, and treatment backgrounds. A clinically interactive segmentation network termed ColSegNet (Colon Cancer Segmentation Network) was developed to incorporate physician guidance into the segmentation workflow. In clinical use, the physician provides quick coarse tumor localization via a rough bounding box or few scribbles as the initial guide. A global semantic prior derived from whole-volume anatomical context is then injected to refine the region of interest and reduce false positives. The model subsequently performs fine-grained delineation within the guided region. Test-time augmentation with Union merging ensures high recall and robust tumor coverage. Segmentation performance was evaluated on internal validation and two external test sets, with clinical utility further assessed in downstream tasks of N-stage and MMR status prediction.Results ColSegNet achieved a DSC of 0.8319 with minimal clinician interaction, representing a 95.42% relative improvement over the previous state-of-the-art fully automatic method (0.4257 DSC). Annotation time was reduced by 95.95%, from over 3,016 seconds for full manual delineation to only 122 seconds per case. In downstream clinical tasks, ColSegNet-derived masks enabled meaningful N-stage prediction (AUC 0.6644) and MMR status prediction (AUC 0.6639), achieving performance comparable to ground truth annotations, whereas CT alone failed in both tasks.Conclusions The proposed interactive ColSegNet framework delivers near-expert-level segmentation accuracy with drastically reduced clinician effort, addressing key barriers to clinical adoption. Its strong performance in downstream tasks underscores its potential to enhance precision oncology workflows, including staging and molecular profiling. Future work will focus on prospective validation and extension to rectal cancer for comprehensive colorectal analysis.Abstract IDDF2026-ABS-0164 Figure 1