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

Colorectal histology classification remains challenging because subtle morphological differences, staining variability, and inappropriate model configurations can reduce generalization. This study proposes COOT-CNN, a lightweight convolutional neural network in which architectural design and training strategy are optimized jointly through the COOT metaheuristic. The search process considers network depth, convolution type, kernel size, feature width, squeeze-and-excitation attention, activation functions, pooling, regularization, loss formulation, learning-rate scheduling, data augmentation, exponential moving average, and mixed-precision training. The selected configuration is retrained independently and evaluated using accuracy, balanced accuracy, macro-F1, ROC-AUC, precision-recall analysis, confidence intervals, and paired statistical testing. Same-split comparisons are conducted against ResNet-50, EfficientNet-B0, and Swin-T under a consistent experimental protocol. COOT-CNN achieved 96.72% test accuracy and 96.71% macro-F1, while providing statistically significant improvements over the evaluated baselines. The optimized model also maintained a compact computational profile, with fewer parameters and lower inference latency than the comparison networks. These findings demonstrate that jointly optimizing architecture and training components can produce an accurate, efficient, and stable model for binary colorectal histology patch classification. The proposed framework should be regarded as a methodological proof-of-concept, with external, patient-level, and whole-slide validation required before clinical application.