Dual-Level Imbalance Mitigation for Single-FoV Colorectal Histopathology Image Classification

Area of Science:

• Digital pathology
• Machine learning for cancer diagnosis
• Histopathology image analysis

Background:

• Single-field-of-view (FoV) histopathological image classification is crucial for colorectal cancer (CRC) diagnosis in resource-limited hospitals.
• Existing methods struggle with severe class imbalance and performance degradation in single-FoV CRC diagnosis.
• Whole-slide imaging (WSI) scanners and extensive storage are often unavailable in mid- to low-tier healthcare facilities.

Purpose of the Study:

• To develop and evaluate a novel dual-level imbalance mitigation (DIM) framework for improved single-FoV CRC histopathology image classification.
• To address the challenges of class imbalance and performance degradation in CRC diagnosis datasets.
• To enhance the diagnostic accuracy of machine learning models in resource-constrained settings.

Main Methods:

• Proposed a dual-level imbalance mitigation (DIM) framework integrating data-level and algorithm-level strategies.
• Employed a global context generative adversarial network (GCGAN) for realistic minority-class image augmentation.
• Introduced a frequency-aware adaptive focal loss (FAFL) to improve class separation.
• Utilized a lightweight receptive field-based convolutional neural network (LRF-CNN) trained with the DIM framework.

Main Results:

• The DIM-equipped LRF-CNN significantly outperformed five state-of-the-art (SOTA) models on a single-FoV colorectal histopathology dataset across multiple metrics.
• Individual components of the DIM framework demonstrated performance enhancements when applied to existing SOTA models.
• The proposed DIM framework showed generalizability and effectiveness across six additional single-FoV histopathological datasets.

Conclusions:

• The dual-level imbalance mitigation (DIM) framework effectively addresses class imbalance and improves classification performance for single-FoV CRC histopathology.
• The proposed GCGAN and FAFL components, integrated with LRF-CNN, offer a robust solution for CRC diagnosis in resource-limited environments.
• The DIM framework presents a promising and generalizable approach for enhancing automated histopathological analysis in digital pathology.