Efficient colorectal cancer classification from histopathological images utilizing lightweight convolutional neural networks.

As one of the most fatal malignancies worldwide, colorectal cancer demands accurate histopathological image analysis to support timely diagnosis and effective clinical decision-making. This study investigates four lightweight convolutional neural network (CNN) variants for binary classification of colon histopathology patches into Colon_Adenocarcinoma and Colon_Benign_Tissue. A unified experimental framework is employed, incorporating standardized preprocessing, data augmentation, class-weighted optimization, and macro-F1-based early stopping to ensure fair architectural comparison. Model behavior is analyzed using training and validation learning curves, confusion matrices, and ROC and precision-recall diagnostics. Across in-domain validation, Lite-V2 consistently achieves the strongest accuracy-efficiency trade-off with a compact parameter footprint and stable convergence. However, evaluation on an independent test split and across external datasets reveals notable performance degradation, highlighting sensitivity to domain shift and underscoring the limitations of lightweight models under distributional variation. These findings emphasize that, while computationally efficient CNNs can support high-throughput screening and decision assistance, robust clinical deployment requires multi-site validation and continued adaptation to heterogeneous histopathological data.