Context fusion network with multi-scale-aware skip connection and twin-split attention for liver tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Manual liver tumor segmentation is time-consuming and labor-intensive for clinicians.
• Automated segmentation is crucial for clinical diagnosis but faces challenges like heterogeneity and fuzzy boundaries.
• Existing methods struggle with the complex nature of tumor tissue.

Purpose of the Study:

• To develop a novel deep learning approach for accurate automated liver tumor segmentation.
• To address challenges of heterogeneity, fuzzy boundaries, and irregularity in tumor tissue.
• To enable 3D tumor counting for improved cancer grading.

Main Methods:

• Proposed a novel deep learning-based approach incorporating a multi-scale-aware (MSA) module and a twin-split attention (TSA) module.
• The MSA module bridges semantic gaps and preserves detailed information.
• The TSA module recalibrates feature map channel responses.

Main Results:

• Achieved a Dice index of 85.97% and Jaccard index of 81.56% on the LiTS2017 dataset, outperforming state-of-the-art methods.
• Demonstrated robustness and generalization with Dice index of 83.67% and Jaccard index of 80.11% on the 3Dircadb dataset.
• Successfully enabled 3D tumor counting from segmentation results for cancer grading.

Conclusions:

• The proposed deep learning method significantly improves automated liver tumor segmentation accuracy.
• The MSA and TSA modules effectively handle tumor heterogeneity and detailed information.
• This approach offers a robust and generalizable solution for clinical diagnosis and cancer grading.