Histopathological image classification plays a crucial role for physicians in computer-aided-diagnosis and biomedical research. The PathMNIST dataset, which is a part of MedMNIST v2 dataset, is derived from colon pathology slides. This dataset is considered as a benchmark dataset for cancer classification. In this study, a comparative performance analysis of three cnovolutional neural network models namely ResNet18, ResNet50, and EfficientNetB0 are applied on PathMNIST dataset. The effectiveness and performance of these models are analyzed using the metrics precision, recall, accuracy, F1-score and area under curver (AUC) under identical experimental conditions. Further, the influcence of Adam and AdamW optimizers on model’s performance is also studied. The results highlight notable differences in prediction, generalization and robustness and emphasizing the tradeoff between light-weight CNN architectures and complex CNN architectures. The findings contribute to deeper understanding of model flexibility for histopathological image classification.