Automatic Deep Learning-based Histopathologic Image Classification

Abstract

Histopathologic image analysis is a critical component in cancer diagnosis, yet traditional manual inspection methods are often time-consuming, subjective, and error-prone. This study presents a fully automated deep learning framework for the classification of histopathologic images stained with Hematoxylin and Eosin (H&E). Leveraging Convolutional Neural Networks (CNNs), particularly DenseNet and ResNet architectures, the proposed system integrates essential components such as data preprocessing, augmentation, hyperparameter optimization, and training automation using the fastai library. Experiments were conducted on the PCam dataset derived from Camelyon16, comprising over 320,000 labeled image patches. Results demonstrate that DenseNet outperforms ResNet in terms of accuracy and AUC, achieving 84.37% test accuracy and 0.96 AUC. The framework shows high reproducibility, efficiency, and potential for clinical deployment, offering a scalable solution to improve diagnostic accuracy in pathology. Future directions include exploring hybrid models and advanced augmentation techniques to further enhance classification performance.