Robust Automated Tumour Segmentation Network Using 3D Direction-Wise Convolution and Transformer.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Neuroscience

Background:

• Semantic segmentation of tumors is vital for cancer diagnosis and treatment planning.
• U-Net and Transformer models show promise but have limitations in voxel-level classification.
• Transformers lack positional encoding and translation equivariance; CNNs lack global features and dynamic attention.

Purpose of the Study:

• To introduce TCTNet, a novel architecture for enhanced 3D medical image segmentation.
• To address limitations of existing Transformer and CNN models in tumor segmentation.
• To improve accuracy and efficiency in cancer diagnosis and treatment planning.

Main Methods:

• Developed TCTNet, featuring a hybrid Transformer-CNN encoder and a 3D Direction-Wise Convolution decoder.
• Utilized the Brain Tumour Segmentation 2021 (BraTS21) dataset for evaluation.
• Tested generalization on two additional datasets from the Medical Segmentation Decathlon.

Main Results:

• TCTNet demonstrated superior performance compared to other 3D segmentation networks on the BraTS21 dataset.
• The proposed architecture showed strong generalization capabilities across multiple tumor datasets.
• An ablation study confirmed the effectiveness of the 3D Direction-Wise Convolution decoder.

Conclusions:

• TCTNet offers a competitive and efficient solution for 3D medical image segmentation.
• The hybrid approach effectively combines the strengths of Transformers and CNNs.
• The model reduces computational effort by 10% while maintaining high segmentation performance.