AI-ready Rectal Cancer MR Imaging: A Workflow for Tumor Detection and Segmentation

Introduction: Magnetic resonance imaging (MRI) is the preferred modality for staging rectal cancer, but accurate tumor segmentation remains challenging due to overlapping tumor and normal tissue appearances, variability in imaging parameters, and subjective interpretation. While AI offers promising solutions for automating segmentation, progress is limited by the lack of high-quality public MRI datasets. The objective of this study was to facilitate collaboration between radiologists and data scientists to improve tumor detection and segmentation on T2-weighted images (T2WIs) streamlining data preparation for the development of AI models.

Methods: A total of 37 patients with rectal cancer were included in this study. Data scientists were trained through radiologist-led sessions, participation in Stanford’s weekly colorectal cancer multidisciplinary tumor board, and review of radiologist annotations and clinical notes in Epic electronic medical records. The data scientists manually segmented rectal tumors on T2WIs, and their segmentations were reviewed and edited by a US board-certified radiologist. Segmentation accuracy was evaluated using the Dice similarity coefficient score (DSC) and Jaccard index (JI).

Results: The data scientists successfully identified rectal tumors in all T2WIs. The data scientists’ segmentations showed strong agreement with radiologist’s edits, achieving a mean DSC of 0.965 [0.939-0.992] and a mean JI of 0.943 [0.900-0.985]. Discrepancies were mainly due to over- or under-segmentation of rectal tumors.

Conclusion: This study demonstrates the feasibility of creating high-quality labeled MRI datasets through collaboration between radiologists and data scientists. Such datasets are critical for training AI models to automate rectal tumor detection and segmentation, potentially improving clinical workflows and efficiency.