Histopathological Image Classification: EfficientNetB3 vs Custom CNN for Cancer Detection

Identification of histopathological images is crucial in the diagnosis of various cancers including the lung and colon cancers. In this study, we evaluate the performance of two deep learning models: Proposes a Lung- Colon CNN hybrid model and the pre-trained EfficientNetB3, on LC25000, with 30,000 images discretized into lung/ colon cancer (15,000), non-cancerous (5,000) images. Images were preprocessed through resizing to All images were resized to 224×224 pixels, normalized, and augmented. The models and were assessed with regards to accuracy, precision, recall, F1-score and time of inference. When the present study applied EfficientNetB3 to the same data set, it obtained improved values of all indices, with accuracy of 94.7%, precision of 93.8%, recall of 94.2%, and F1-score of 94.0% compared to 88.5% accuracy, 87.2% precision, 86.9% recall, and F1-score of 87.0% of the In addition, EfficientNetB3 achieved shorter inference time of 8.9ms per image for non-real time diagnostic refinement and the custom CNN 12.4ms. Higher efficiency of EfficientNetB3 further reflected by ROC curve analysis wherein the model obtained AUC values closely to 1, meaning better ability to discriminate between classes as opposed to lower AUC values in customized CNN. The Grad-Cam based visualization showed that EfficientNetB3 is selective to areas in histopathological slides making it easier to interpret. It is clear from the analysis that a family of efficient models has significant architectural benefits including compound scaling and effective implementation of squeeze and excitation blocks that gives EfficientNetB3 a better performance with less computation.