Optimized Histopathological Image Classification for Breast Cancer Using Deep Learning Model

Abstract

A significant cause of death very early among women is due to breast cancer. Since invasive ductal carcinoma (IDC) and breast cancer in general continue to be among the most common and fatal illnesses affecting women, prompt identification is crucial. Convolutional neural networks (CNNs), in particular, are particularly notable in automating image processing of breast cancer to the point that the images do not need human interpretation. The proposed project enhances the detection of IDC by creating a reliable diagnostic algorithm, in which deep learning and histopathology image analysis are used. On a large-scale dataset of IDCs containing more than 277,000 image patches, it applied a full preprocessing pipeline, including Otsu thresholding, tissue masking, Hematoxylin channel extraction, CLAHE enhancement, Gaussian denoising and gamma correction to improve image quality. Images with low tissue content were eliminated and SMOTE has been utilized to work on the imbalance of classes. This was a fully fine-tuned ResNet50V2 that was pretrained on ImageNet, and then combined with own dense layers and trained based on Adam optimizer and binary classification. Accuracy (acc) of the model was 88.52%, and precision (pre), recall (rec), and F1-score (F1) were greater than 88%. The effect of various architectural and training setups was studied using ablation and confirmed the efficiency of the chosen model. The comparative analysis yielded better performance in comparison to the existing CNN, CNN-GRU, and DenseNet-based models. The results indicate the possibility of AI-based breast cancer detection to be used in clinical practice, decrease errors in diagnosis, and increase the rate of early diagnoses. Combined application of the state-of-the-art image enhancement, balanced data depiction, and comprehensive study of the ablation using pre-trained CNN models is novel. This adds a high-performing and interpretable, yet clinically relevant, solution to the early diagnosis of breast cancer by using histopathology.