ASU-Net++: A nested U-Net with adaptive feature extractions for liver tumor segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Deep Learning for Diagnostics

Background:

• Accurate tumor localization is crucial for medical diagnosis and treatment planning.
• Segmenting tumors with complex shapes presents a significant challenge in medical image analysis.
• Existing methods often struggle with diverse tumor sizes and intricate boundaries.

Purpose of the Study:

• To develop an advanced deep learning model for automated tumor segmentation in medical images.
• To enhance feature extraction and preservation for improved segmentation accuracy.
• To validate the model's performance on ultrasound and CT datasets with varying tumor characteristics.

Main Methods:

• A multi-level feature extraction neural network employing a modified nested U-Net architecture.
• Integration of dilated dense short skip connections for improved gradient flow and feature preservation.
• Implementation of an adaptive Atrous Spatial Pyramid Pooling (ASPP) module for multi-level feature extraction.
• Utilized the AdaBelief optimizer for faster convergence.

Main Results:

• The proposed model achieved high segmentation accuracy across liver tumor ultrasound and CT datasets.
• Demonstrated superior performance compared to multiple existing network structures.
• Achieved dice coefficients of 0.9153 (ultrasound), 0.9413, and 0.9246 (CT datasets).
• Showcased effective generalization capabilities with small and diverse datasets.

Conclusions:

• The novel deep learning model offers a robust solution for accurate tumor segmentation in medical imaging.
• The adaptive ASPP and enhanced U-Net architecture effectively handle complex tumor morphologies.
• The model's performance indicates significant potential for improving diagnostic accuracy and clinical workflows.