Accurate colorectal tumor segmentation for CT scans based on the label assignment generative adversarial network

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Manual segmentation of colorectal tumors in CT scans is time-consuming and prone to variability.
• Deep learning networks offer automated segmentation but often require refinement for optimal accuracy.
• Accurate segmentation is crucial for effective colorectal cancer diagnosis and treatment planning.

Purpose of the Study:

• To develop and evaluate an automatic post-processing module, the Label Assignment Generative Adversarial Network (LAGAN), for refining deep network-based colorectal tumor segmentation.
• To assess the performance of LAGAN when applied to segment colorectal tumors in computed tomography (CT) scans.
• To explore the efficacy of LAGAN in conjunction with different deep network architectures (FCN32 and Unet).

Main Methods:

• The study enrolled 223 patients with colorectal cancer (CRC).
• CT scans were initially segmented using Fully Convolutional Network 32 (FCN32) and Unet, generating probabilistic maps.
• The LAGAN, comprising a generative and a discriminative model, refined these probabilistic maps to achieve final binary segmentation of colorectal tumors.

Main Results:

• LAGAN improved the Dice Similarity Coefficient (DSC) for FCN32 from 81.83% to 90.82%.
• For Unet-based segmentation, LAGAN increased the DSC from 86.67% to 91.54%.
• The refinement process using LAGAN was rapid, taking approximately 10 milliseconds per CT slice.

Conclusions:

• The LAGAN module demonstrates robustness and flexibility in refining segmentation from various deep networks.
• LAGAN achieves high segmentation accuracy for colorectal tumors, outperforming other methods.
• This automated approach offers a significant advancement for efficient and accurate colorectal tumor segmentation in clinical practice.