AMC-Net: Asymmetric and multi-scale convolutional neural network for multi-label HPA classification

Area of Science:

• Biomedical image analysis
• Protein localization
• Computational biology

Background:

• Accurate Human Protein Atlas (HPA) classification aids disease understanding and biomedical image analysis.
• Existing automatic protein recognition methods are limited to single pattern recognition.
• A need exists for an automatic multi-label HPA recognition system with enhanced performance.

Purpose of the Study:

• To develop an automatic recognition system for multi-label HPA classification using deep learning.
• To address the limitations of single-pattern recognition in current protein identification methods.

Main Methods:

• Proposed an automatic feature extraction and multi-label classification framework.
• Designed an asymmetric and multi-scale convolutional neural network for HPA classification.
• Introduced a combined loss function (binary cross-entropy and F1-score) to enhance identification performance.

Main Results:

• The system successfully classifies mixed protein patterns in microscope images.
• The framework handles multi-label subcellular protein classification across 28 patterns.
• Achieved a F1-score of 0.823, outperforming existing methods and demonstrating robustness.

Conclusions:

• A high-performance deep learning system for protein atlas classification was developed.
• The study presents an automatic multi-label HPA identification framework with superior performance.
• The proposed system enhances the automatic analysis of biomedical images for protein localization.