IMAU-Net: A Hybrid Multi-Scale Deep Learning Framework for Liver Segmentation from Laparoscopic Images

Area of Science:

• Medical Image Analysis
• Surgical Technology
• Artificial Intelligence in Medicine

Background:

• Accurate liver segmentation is crucial in laparoscopic surgery but challenging due to visual complexities.
• Existing deep learning models struggle to balance segmentation accuracy, computational efficiency, and boundary precision.

Purpose of the Study:

• To develop and evaluate IMAU-Net, a novel hybrid deep learning architecture for precise and efficient liver segmentation in laparoscopic surgery.
• To address limitations of current models by integrating advanced feature extraction techniques.

Main Methods:

• Proposed IMAU-Net architecture: a hybrid model combining a pre-trained InceptionV3 encoder with a novel bottleneck.
• The bottleneck integrates Multi-Core Pooling (MCP) for fine-to-medium spatial details and enhanced Atrous Spatial Pyramid Pooling (ASPP) for multi-scale context.
• Evaluated using 5-fold cross-validation on the M2CAI dataset and external validation on the CholecSeg8K dataset.

Main Results:

• IMAU-Net achieved a mean Dice coefficient of 0.9179 ± 0.012 and IoU of 0.8483 ± 0.015 on the M2CAI dataset.
• External validation on CholecSeg8K dataset yielded a Dice coefficient of 0.8745 ± 0.0312 and AUC of 0.9542, showing generalizability.
• The model demonstrates superior performance compared to state-of-the-art methods with high computational efficiency (45 FPS, 42.3 M parameters).

Conclusions:

• IMAU-Net offers an optimal balance between accuracy and efficiency for liver segmentation in laparoscopic surgery.
• The model shows potential for integration into real-time intraoperative guidance systems.
• Further prospective validation is recommended for real-time intraoperative workflow integration.