A deep learning strategy for the 3D segmentation of colorectal tumors from ultrasound imaging

Colorectal cancer remains a leading cause of cancer-related mortality worldwide, highlighting the need for accurate and efficient diagnostic tools. While Deep Learning has shown promise in medical imaging, its application to transrectal ultrasound for colorectal tumor segmentation remains underexplored. Currently, lesion segmentation is performed manually, relying on clinician expertise and leading to significant variability across treatment centers. To overcome this limitations, we propose a novel strategy that addresses both practical challenges and technical constraints, particularly in scenarios with limited data availability, offering a robust framework for accurate 3D colorectal tumor segmentation from ultrasound imaging. We evaluate eight state-of-the-art models, including convolutional neural networks and transformer-based architectures, and introduce domain-tailored pre- and post-processing techniques such as data augmentation, patching and ensembling to enhance segmentation performance while reducing computational cost. Leading to an average improvement in term of DICE score of 0.423 absolute points (+107%), compared to baseline models, our findings demonstrate the potential of our proposal to improve the accuracy and reliability of ultrasound-based diagnostics for colorectal cancer, paving the way for clinically viable AI-driven solutions.