The Automatic Segmentation of Rectal Cancer on MRIs using nnU-Net

Introduction: Segmentations are required for MRI-based radiomic ML models to predict response to treatment. However, manual segmentations of rectal tumors by radiologists are laborious, time-consuming, and costly. Moreover, in the case of rectal cancer, there are no publicly available MRI datasets or models. To address these challenges, we trained an automatic 3D segmentation model, No New U-Net (nnU-Net) on rectal MRIs and tumor segmentations. 

Methods: We utilized MRI data from 37 patients enrolled in the SHORT-FOX rectal cancer treatment trial. T2-weighted axial smFOV MRIs were manually segmented using 3D Slicer v5.4.0 by an expert radiologist. We used a 5-fold cross validation to train and assess the nnU-Net’s performance in segmenting rectal tumors on MRIs. Model performance was assessed using Dice Similarity Coefficient (DSC), which ranges from 0, indicating no overlap, and 1, indicating perfect overlap between manual and nnU-Net segmentations.

Results: The nnU-Net successfully detected all 37 rectal tumors correctly with a median DSC (IQR) of 0.7 (0.42-0.74). The largest discrepancies between the manual and nnU-Net segmentations can be attributed to the inclusion of additional slices beyond manual segmentation boundaries and the presence of adjacent organs or intraluminal stool or gas.

Conclusion: Our study highlights the potential of the nnU-Net model to enhance accessibility to advanced imaging analysis for assessing tumor features and treatment response among rectal cancer patients. Our nnU-Net model trained on rectal cancer MRIs will be made publicly available for use and refinement.