A Holistic Strategy of Modified Superpixel Segmentation and Randomized Adam Hyperparameter Tuning with Deep Learning Approaches for the Classification of Breast Cancer from BreakHis Images: In the Quest for Precision

Breast cancer is a prevalent cancer type in women worldwide, and therefore it is necessary to do early detection that is accurate for effective treatment. However, traditional ways of diagnosing through mammogram or histopathological examination may take more time and also require an interpretation from an expert. In the past few years, deep learning techniques especially Convolutional Neural Networks (CNNs) have changed medical imaging by making possible automated diagnosis which is fast. In this study, an integrated breast cancer detection approach using BreakHis images is proposed focusing on balanced accuracy rate analysis to solve imbalanced datasets problem. The methodology starts with pre-processing these images by means of Adaptive Fuzzy Filter which removes the artifacts while improving the image quality. The next step involves superpixel segmentation through SLIC-K-Means-BIRCH followed by feature extraction using SIFT with Bag of Features (BoF). The extracted features are analyzed with machine learning models such as Gaussian Mixture Model (GMM), Decision Tree (DT), Softmax Discriminant Classifier (SDC), SVM using RBF kernel and Naive Bayes Classifier (NBC) as well as deep learning models ResNet-50, VGG16, VGG19 and EfficientNet-B0. Data augmentation techniques such as image rotation and brightness adjustments were applied to enrich our dataset. R-Adam hyperparameter tuning technique was utilized to optimize these deep learning models. Results show that the modified SLIC-K-Means-BIRCH segmentation, when combined with the SIFT with BoF and EfficientNet-B0 optimized with R-Adam, yields a classification accuracy of 99.11% on augmented images. In addition, balanced classification rate analysis confirmed that this method could classify breast cancer from an imbalanced dataset effectively.