3-D RoI-Aware U-Net for Accurate and Efficient Colorectal Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Accurate segmentation of colorectal cancerous regions in 3-D MR images is vital for radiotherapy.
• Existing deep learning methods face challenges with GPU memory, effective receptive field size, and computational cost for 3-D whole volume segmentation.

Purpose of the Study:

• To develop a novel, efficient, and accurate deep learning framework for 3-D whole volume segmentation of colorectal cancer.
• To address the limitations of existing methods in terms of memory footprint, speed, and segmentation performance.

Main Methods:

• Proposed a 3-D RoI-aware U-Net (3-D RU-Net) framework with a global image encoder for RoI localization and a local region decoder operating on pyramid-shaped features.
• Implemented a dice-based multitask hybrid loss function to enhance global-to-local learning and contour detail.
• Enabled training and prediction on large 3-D whole volumes due to GPU memory efficiency.

Main Results:

• Achieved a 75.5% Dice Similarity Coefficient (DSC) with a processing time of 0.61 seconds per volume on a GPU.
• Significantly outperformed competing methods in both accuracy and efficiency on a dataset of 64 T2-weighted MR images.
• Demonstrated the effectiveness of model ensembling for further performance gains.

Conclusions:

• The 3-D RU-Net framework offers a significant advancement in automatic 3-D whole volume segmentation for colorectal cancer.
• The method provides a computationally efficient and accurate solution for radiotherapy applications.
• The publicly available code facilitates further research and development in medical image segmentation.