Fully semantic segmentation for rectal cancer based on post-nCRT MRl modality and deep learning framework.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiomics

Background:

• Accurate rectal tumor segmentation on post-neoadjuvant chemoradiotherapy (nCRT) MRI is crucial for treatment evaluation and surgical planning.
• Manual segmentation is time-consuming and subject to inter-observer variability.
• Deep learning offers a potential solution to automate and standardize this process.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automated rectal tumor segmentation.
• To assess the segmentation performance exclusively on post-chemoradiation T2-weighted MRI.
• To reduce the workload for radiologists and clinicians in detecting and measuring rectal tumors.

Main Methods:

• Retrospective analysis of 372 LARC patients treated with standard nCRT.
• Development of a symmetric eight-layer deep network using the nnU-Net Framework.
• Training and validation on 243 patients (3061 slices) and testing on 41 patients (408 slices) using fivefold cross-validation.

Main Results:

• The deep learning model achieved an average Dice Similarity Coefficient (DSC) of 0.700 on the test dataset.
• Quantitative evaluation showed an average 95% Hausdorff Distance (HD95) of 17.73 mm and Mean Surface Distance (MSD) of 3.11 mm.
• A significant portion of MSD values were below 5 mm (82%) and 2 mm (55%), indicating high accuracy for most cases.

Conclusions:

• The developed deep learning pipeline demonstrates relatively high accuracy for rectal tumor segmentation on post-nCRT MRI.
• The findings suggest the potential of AI to assist in clinical workflows for rectal cancer management.
• Future research should focus on multicenter external validation to confirm generalizability.