A fully automatic deep learning algorithm to segment rectal Cancer on MR images: a multi-center study.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate segmentation of colorectal cancer (CRC) on MRI is crucial for diagnosis and treatment planning.
• Manual segmentation is time-consuming and subject to inter-observer variability.
• Deep learning offers potential for automated and standardized segmentation.

Purpose of the Study:

• To develop and optimize a fully automatic deep learning algorithm for CRC segmentation on MR images.
• To evaluate the algorithm's performance across multiple institutions and MRI scanners.
• To establish a reliable tool for subsequent radiomics analyses in CRC management.

Main Methods:

• A U-Net based deep learning architecture was employed for CRC segmentation.
• The algorithm included a pre-processing step for image normalization and tumor highlighting.
• Performance was assessed by comparing automated segmentations against manual segmentations by experienced radiologists.

Main Results:

• The best performing models (mdl2, mdl3) achieved median Dice Similarity Coefficients of 0.68 and 0.69, respectively.
• High detection rates of 0.98 (mdl2) and 0.95 (mdl3) were observed on the validation set.
• The algorithm demonstrated robust performance across different MRI scanners and institutions.

Conclusions:

• The developed deep learning algorithm provides a fast and precise tool for automatic CRC segmentation on MR images.
• This automated approach has the potential to enhance the management of colorectal cancer patients.
• The tool can serve as a foundational step for future radiomics studies predicting treatment response and personalizing therapy.