•  
  •  
 

Abstract

Pathological anatomical images play a pivotal role in diagnosing diseases, notably breast cancer, which affects women globally. These images, obtained through biopsies or post-mortem examinations, are preserved to maintain their structural integrity. Software tools, like computer-aided diagnosis, aid doctors in early detection and treatment planning, contributing to reduced mortality rates. In this context, convolutional neural networks (CNNs) have emerged as valuable tools for diagnosing benign and malignant breast cancers. This paper introduces a Mega Ensemble Net method, leveraging multi-scale combination features on the breast histopathology dataset. Three fine-tuned deep learning models, namely ResNet-18, ResNet-34, and ResNet-50, are integrated into this method. Techniques such as patch extraction for data augmentation, dataset amalgamation, and transfer learning bolster the method's capabilities. Fusing extracted patches with primary images enhances the method's robustness and adaptability, offering diverse perspectives and intricate details for nuanced class distinctions. BACH and BreaKHis datasets have been used to evaluate the Mega Net. During four‑fold cross‑validation on the test folds, the Mega-Net demonstrates 99% test‑set accuracy in the full image and 98% test‑set accuracy in patches within the multi-classification BACH dataset and 99% test‑set accuracy within the binary classification BreaKHis dataset. Moreover, the MEMF-Net achieved a multi-classification test accuracy of 98.95% across an optimal selected MEMF model in validation testing images.

Share

COinS