Transfer Learning Approach with Features Block Selection via Genetic Algorithm for High-Imbalance and Multi-Label...

Area of Science:

• Computational biology
• Medical imaging analysis
• Deep learning applications

Background:

• Deep learning excels in image classification but faces challenges in complex medical imaging datasets like the Human Protein Atlas (HPA).
• The HPA dataset presents computational difficulties due to high class imbalance, rare patterns, and multi-label classification requirements.
• The dataset includes 28 patterns, over 500 label combinations, and four distinct imaging channels (green, red, blue, yellow) per sample.

Purpose of the Study:

• To develop an effective deep learning strategy for analyzing the complex Human Protein Atlas dataset.
• To improve the accurate identification of rare protein localization patterns within cellular structures.
• To leverage transfer learning and feature extraction from pre-trained Convolutional Neural Networks (CNNs) for enhanced classification.

Main Methods:

• A two-phase transfer learning approach using feature-block extraction from twelve ImageNet-pretrained CNNs.
• Phase one: Single-label multiclass classification with CNNs as feature extractors and Support Vector Machine (SVM) classifiers, incorporating a genetic algorithm for feature block selection.
• Phase two: Application of two multi-label classification strategies, integrating image-level (channel analysis) and cell-level (nucleus and nuclear-membrane ring features) analyses.

Main Results:

• Concatenating feature blocks from different CNNs significantly improved performance in the first phase.
• Image-level analysis revealed the green channel's individual strength and the benefit of combining it with red and yellow channels.
• Cell-level analysis focusing on the nucleus and nuclear-membrane ring effectively recognized rare patterns, outperforming existing methods.

Conclusions:

• The proposed two-phase transfer learning method effectively addresses the challenges of the HPA dataset, particularly rare pattern identification.
• Integration of image-level and cell-level analyses enhances the detection of subtle and infrequent protein localizations.
• The study achieved a notable F1 score of 0.8 for the "Rods & Rings" pattern, demonstrating significant progress in biological and clinical applications.