COOT-CNN: joint architecture and training-strategy optimization for binary colorectal histology patch classification

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for histology

Background:

• Colorectal histology classification is complex due to subtle morphological variations and staining inconsistencies.
• Existing deep learning models can struggle with generalization in histology tasks.
• Optimizing both model architecture and training is crucial for robust performance.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) for accurate and efficient colorectal histology patch classification.
• To jointly optimize the CNN architecture and training strategy using the COOT metaheuristic.
• To evaluate the performance of the proposed model against established benchmarks.

Main Methods:

• The COOT metaheuristic was employed to jointly optimize CNN architecture (depth, convolution type, kernel size, etc.) and training strategy (loss formulation, learning rate, data augmentation, etc.).
• The selected COOT-CNN configuration was retrained and evaluated using metrics including accuracy, macro-F1, ROC-AUC, and statistical testing.
• Comparative analysis was performed against ResNet-50, EfficientNet-B0, and Swin-T using a consistent experimental protocol.

Main Results:

• COOT-CNN achieved a test accuracy of 96.72% and a macro-F1 score of 96.71%.
• The model demonstrated statistically significant improvements over ResNet-50, EfficientNet-B0, and Swin-T.
• COOT-CNN exhibited a more compact computational profile with fewer parameters and lower inference latency compared to baselines.

Conclusions:

• Joint optimization of architecture and training components yields an accurate, efficient, and stable model for binary colorectal histology classification.
• The proposed COOT-CNN framework offers a promising methodological approach.
• Further external, patient-level, and whole-slide validation is necessary before clinical implementation.