AMCOD: Acute Malignant Colorectal Obstruction Dataset

Automated colorectal segmentation provides a quantitative basis for the preoperative assessment of acute malignant colorectal obstruction. However, existing public datasets predominantly focus on conventional anatomical structures, lacking representation of colorectal cancer-related anatomical phenotypes. This study aims to establish a dedicated dataset for colorectal segmentation in acute malignant colorectal obstruction and provide a benchmark for evaluating segmentation models.

A total of 86 eligible cases were retrospectively included to construct AMCOD, with the colorectum annotated into proximal, obstructed, and distal segments. Experiments were conducted to establish a segmentation benchmark on AMCOD, assess the contribution of the three-region supervision strategy through ablation analysis, and evaluate cross-dataset transferability with two public colorectal datasets.

Benchmark experiments revealed differences across the three annotated colorectal regions, with the obstructed segment consistently showing the lowest segmentation performance across the evaluated models. Ablation analysis showed that three-region supervision improved obstructed-segment segmentation compared with obstruction-only supervision. Cross-dataset evaluation further showed that the AMCOD-trained model retained colorectal segmentation capability on both evaluated external datasets.

This work provides a dataset resource for acute malignant colorectal obstruction segmentation and establishes a benchmark for evaluating segmentation methods under obstruction-related anatomical conditions. To facilitate further research and community development, the proposed dataset is made publicly available at https://github.com/FZU312lab/AMCOD.