Pathologic liver tumor detection using feature aligned multi-scale convolutional network.

Area of Science:

• Digital pathology
• Computational oncology
• Medical image analysis

Background:

• Accurate detection of hepatocellular carcinoma (HCC), the most common liver tumor, is crucial for liver pathology.
• Similarities between tumor and benign tissue cell changes (apoptosis, necrosis, steatosis) can hinder HCC detection.
• Limited tissue regions in image patches may lack sufficient neighboring cell structure information, leading to detection failures.

Purpose of the Study:

• To propose an automated liver tumor detection method using whole slide images (WSI).
• To enhance HCC detection by integrating features from multiple magnification levels for improved contextual information.

Main Methods:

• Development of a Feature Aligned Multi-Scale Convolutional Network (FA-MSCN) architecture.
• Utilizing two parallel convolutional networks: one for high-resolution features, another for low-resolution features via atrous convolution.
• Integrating low-resolution features with high-resolution features through central cropping, upsampling, and concatenation for classification.

Main Results:

• Multi-Scale Convolutional Network (MSCN) demonstrated improved detection performance over Single-Scale Convolutional Network (SSCN).
• The proposed FA-MSCN significantly outperformed both SSCN and MSCN in HCC detection.
• The FA-MSCN effectively leveraged neighboring cell structure information for more robust detection.

Conclusions:

• The FA-MSCN architecture offers a superior approach for automated HCC detection in WSIs.
• Integrating multi-scale features is key to overcoming limitations of single-scale analysis in liver pathology.
• This method holds promise for improving the accuracy and efficiency of liver tumor diagnosis.