Data-Oriented Octree Inverse Hierarchical Order Aggregation Hybrid Transformer-CNN for 3D Medical Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Hybrid CNN-transformer models advance medical image segmentation by combining global and local feature extraction.
• Current research often overlooks data structure optimization, limiting performance, efficiency, and generalization.

Purpose of the Study:

• To introduce nnU-OctTN, a data-oriented hybrid CNN-transformer model for 3D medical image segmentation.
• To leverage an octree data structure for improved segmentation performance, resource efficiency, and model interpretability.

Main Methods:

• Developed nnU-OctTN, a U-Net-based framework with a node aggregation transformer encoder and an octree data structure.
• Implemented a cross-fusion module for multi-resolution feature map learning, connecting encoder and decoder.
• Utilized dataset-specific parameter adaptation, inspired by nnUNet, avoiding pre-trained weights.

Main Results:

• Achieved high Dice Score Coefficients (DSC) on BTCV (86.95%), ACDC (92.82%), and BraTS (90.61%) datasets.
• Demonstrated generalizability and effectiveness across diverse medical imaging datasets.
• Ablation studies validated the cross-fusion module's effectiveness and the model's scalability.

Conclusions:

• nnU-OctTN offers a promising approach for high-quality 3D medical image segmentation.
• The data-oriented strategy, using octree structures, enhances performance and generalizability.
• The framework presents a competitive and effective solution for clinical applications.