AI-ready rectal cancer MR imaging: a workflow for tumor detection and segmentation.

Abstract

Background: Magnetic Resonance (MR) imaging is the preferred modality for staging in rectal cancer; however, despite its exceptional soft tissue contrast, segmenting rectal tumors on MR images remains challenging due to the overlapping appearance of tumor and normal tissues, variability in imaging parameters, and the inherent subjectivity of reader interpretation. For studies requiring accurate segmentation, reviews by multiple independent radiologists remain the gold standard, albeit at a substantial cost. The emergence of Artificial Intelligence (AI) offers promising solutions to semi- or fully-automatic segmentation, but the lack of publicly available, high-quality MR imaging datasets for rectal cancer remains a significant barrier to developing robust AI models.

Objective: This study aimed to foster collaboration between a radiologist and two data scientists in the detection and segmentation of rectal tumors on T2- and diffusion-weighted MR images. By combining the radiologist's clinical expertise with the data scientists' imaging analysis skills, we sought to establish a foundation for future AI-driven approaches that streamline rectal tumor detection and segmentation, and optimize workflow efficiency.

Methods: A total of 37 patients with rectal cancer were included in this study. Through radiologist-led training, attendance at Stanford's weekly Colorectal Cancer Multidisciplinary Tumor Board (CRC MDTB), and the use of radiologist annotations and clinical notes in Epic Electronic Health Records (EHR), data scientists learned how to detect and manually segment tumors on T2- and diffusion-weighted pre-treatment MR images. These segmentations were then reviewed and edited by a radiologist. The accuracy of the segmentations was evaluated using the Dice Similarity Coefficient (DSC) and Jaccard Index (JI), quantifying the overlap between the segmentations delineated by the data scientists and those edited by the radiologist.

Results: With the help of radiologist annotations and radiology notes in Epic EHR, the data scientists successfully identified rectal tumors in Slicer v5.7.0 across all evaluated T2- and diffusion-weighted MR images. Through radiologist-led training and participation at Stanford's weekly CRC MDTB, the data scientists' rectal tumor segmentations exhibited strong agreement with the radiologist's edits, achieving a mean DSC [95% CI] of 0.965 [0.939-0.992] and a mean JI [95% CI] of 0.943 [0.900, 0.985]. Discrepancies in segmentations were attributed to over- or under-segmentation, often incorporating surrounding structures such as the rectal wall and lumen.

Conclusion: This study demonstrates the feasibility of generating high-quality labeled MR datasets through collaboration between a radiologist and two data scientists, which is essential for training AI models to automate tumor detection and segmentation in rectal cancer. By integrating expertise from radiology and data science, this approach has the potential to enhance AI model performance and transform clinical workflows in the future.