Abdomen CT multi-organ segmentation using token-based MLP-Mixer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy Planning

Background:

• Manual organ contouring in radiotherapy is time-consuming and prone to variability.
• Automated segmentation using deep learning is crucial for efficient and accurate treatment planning.

Purpose of the Study:

• To investigate an efficient deep-learning-based segmentation algorithm for abdomen CT images.
• To facilitate radiation treatment planning through improved segmentation.

Main Methods:

• A novel deep learning model, MLP-Vnet, combining U-shaped MLP-Mixer and CNNs for multi-organ segmentation in abdomen CT.
• The model utilizes MLP-Convolutional blocks for local and global feature extraction, and pixel-level detail recovery.
• Performance evaluated on institutional (60 cases) and public (BCTV, 30 cases) datasets using Dice score, Hausdorff distance, and computational complexity.

Main Results:

• MLP-Vnet achieved high accuracy on the institutional dataset (DSC=0.912) and public dataset (DSC=0.786).
• Demonstrated statistically significant improvements in segmentation accuracy for specific organs (lung, spinal cord, stomach, pancreas, adrenal glands) compared to state-of-the-art networks.
• Generated contours in under 5 seconds, showing significantly faster inference times and lower memory complexity.

Conclusions:

• The proposed MLP-Vnet offers superior segmentation accuracy and efficiency compared to existing methods.
• This reliable and efficient approach has the potential to optimize clinical workflows in abdominal radiotherapy, particularly for online adaptive treatments.