Breast Cancer Classification Using Densenet121 And K-Means Segmentation With Augmented Data

Abstract

Breast cancer remains a significant global health challenge, necessitating improved diagnostic approaches for early detection and treatment. This study presents an optimized deep learning framework that integrates DenseNet121 with K-Means clustering for enhanced segmentation and feature extraction in breast cancer histopathology images. The BreakHis dataset, comprising 7,909 images at varying magnifications (40×, 100×, 200×, and 400×), was employed for model training and evaluation. Image preprocessing involved histogram equalization and augmentation techniques, including rotation and contrast adjustment, to enhance model robustness. The DenseNet121 model was fine-tuned using transfer learning with pre-trained ImageNet weights, and hyperparameters were optimized to improve classification performance. The proposed model achieved an accuracy of 95.21%, surpassing conventional architectures such as ResNet50 (92.4%) and Xception (88.08%). Additionally, an external validation on the BACH dataset demonstrated an accuracy of 92.10%, reinforcing the model's generalizability. Comparative analysis and ablation studies confirmed the significance of K-Means clustering in improving classification outcomes. Future research will focus on multi-modal imaging techniques and Explainable AI (XAI) to enhance interpretability and clinical applicability.