Rethinking Architectural Complexity in Deep Vision Models for Histopathological Image Classification

Area of Science:

• Deep learning applications in digital pathology
• Computational efficiency in medical imaging analysis
• Histopathological image classification

Background:

• Deep learning models for histopathological analysis are becoming more complex, demanding significant computational resources and large datasets.
• Despite high performance, complex models often do not translate to practical clinical benefits for practitioners.

Purpose of the Study:

• To systematically analyze performance-efficiency trade-offs between CNN and transformer architectures in histopathological image classification.
• To evaluate model performance under different data availability scenarios.
• To provide empirical recommendations for selecting appropriate deep learning architectures based on data, balance, and computational constraints.

Main Methods:

• Comprehensive evaluation of popular CNN and transformer models on three distinct histopathological tissue classification datasets.
• Assessment of model performance using standard and clinically relevant diagnostic metrics.
• Analysis of training time and computational resource requirements.

Main Results:

• EfficientNet-B0 and ResNet-50 achieved high performance (F1 ≈ 0.983 at 10% data) with 2-5x faster training compared to transformers.
• Transformers showed modest gains, primarily on imbalanced datasets (F1 ≈ 0.865-0.964), but required substantially higher computational resources.
• Statistical analyses confirmed the significance of observed performance and efficiency trends.

Conclusions:

• Simpler CNN architectures (EfficientNet-B0, ResNet-50) offer a superior balance of performance and efficiency for histopathological image classification, especially with limited data.
• Recommendations are provided for practical architecture selection in resource-constrained clinical settings.
• Findings require validation in real-world clinical environments.