An improved deep learning-based MSA-Net model for small liver tumor segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Radiology

Background:

• Liver cancer is a leading cause of cancer mortality globally, necessitating early detection for improved patient outcomes.
• Accurate segmentation of small liver tumors in complex medical images remains a challenge for conventional deep learning models like U-Net.
• Existing methods struggle to differentiate small or irregularly shaped tumors within intricate liver structures.

Purpose of the Study:

• To develop an automated computer-aided diagnosis (CAD) system for high-precision small liver tumor segmentation.
• To enhance radiologists' capabilities in identifying and delineating small liver tumors in medical imaging.
• To address the limitations of current deep learning techniques in segmenting small and complex-shaped tumors.

Main Methods:

• Proposed MSA-Net (Multi-Scale Attention Network), a U-Net variant incorporating multi-scale convolutional layers and an attention mechanism.
• The architecture integrates multi-scale feature extraction in both encoder and decoder pathways.
• An attention mechanism focuses on salient regions and aggregates contextual information across multiple receptive fields.

Main Results:

• MSA-Net demonstrated superior performance on the 3DIRCADb and LiTS datasets.
• On the 3DIRCADb dataset, MSA-Net achieved a Dice score of 92.00% and a Jaccard index of 86.00% for small tumors.
• On the LiTS dataset, MSA-Net achieved a Dice score of 72.57% and a Jaccard index of 65.35% for small tumors.

Conclusions:

• MSA-Net significantly improves tumor segmentation accuracy, especially for small and challenging liver tumors.
• The study's independent evaluation of large and small tumors highlights MSA-Net's specific advantage in detecting smaller lesions.
• The developed MSA-Net shows strong potential for clinical application in liver cancer diagnosis and treatment planning.