Transformers for colorectal cancer segmentation in CT imaging.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Radiology and oncology

Background:

• Transformer models are emerging as state-of-the-art in medical image segmentation.
• Conventional deep learning approaches, like convolutional neural networks (CNNs), have been the standard for such tasks.
• Colorectal cancer (CRC) segmentation in CT imaging presents a significant challenge.

Purpose of the Study:

• To apply various transformer models to CRC segmentation in CT imaging.
• To compare transformer model performance against the current state-of-the-art CNN, nnUnet.
• To investigate the impact of network size on transformer model accuracy for CRC segmentation.

Main Methods:

• Implementation of six transformer models with varying architectures and sizes.
• Application of these models to the CRC segmentation task within the Medical Segmentation Decathlon.
• Comparison of transformer model performance with nnUnet and inter-observer variability (IOV).

Main Results:

• Swin-UNETR, D-Former, and VT-Unet achieved the highest Dice similarity coefficients (DSC), reaching 0.60, 0.59, and 0.59, respectively.
• Transformer architectures outperformed the nnUnet for CRC segmentation.
• Achieved DSC scores were comparable to inter-observer variability (approx. 0.64), indicating near expert-level accuracy.

Conclusions:

• Transformer models demonstrate state-of-the-art performance in challenging CRC segmentation tasks.
• While accuracies are improving, advances are becoming marginal, suggesting other factors like efficiency are increasingly important.
• The study highlights the potential of transformers but also the inherent challenges and limitations in achieving higher segmentation accuracy due to task complexity and IOV.