GWO-Tuned Deep Learning Model with Memory Mechanism for Accurate Mammographic Breast Cancer Diagnosis

Resumen

Automated and accurate diagnosis of breast cancer from mammography images remains a critical challenge in medical imaging, necessitating advanced computational approaches to improve early detection rates. This paper proposes a hybrid deep learning framework designed to enhance diagnostic precision of breast cancer. The methodology commences with a meticulous preprocessing pipeline, including Contrast Limited Adaptive Histogram Equalization (CLAHE) with optimized Rayleigh distribution and clip limits, a novel background cropping technique to focus on relevant tissue, pixel intensity adjustments for zero-valued regions, and data augmentation through rotations, flips, and zooming. Feature extraction is subsequently performed using a pre-trained ResNet-50 architecture, adapted via transfer learning with fine-tuning of its terminal layers and a custom dense layer. The resultant high-dimensional feature vectors are then refined using Linear Discriminant Analysis (LDA) to enhance class separability while reducing dimensionality. For classification, a Bidirectional Gated Recurrent Unit (BiGRU) network is employed, with its crucial hyperparameters (number of hidden units, learning rate, decay factor, and batch size) systematically optimized using the Grey Wolf Optimizer (GWO). The developed model demonstrates strong diagnostic performance, achieving an accuracy of 98.44% on the INbreast dataset, highlighting its potential as an effective tool for computer-aided breast cancer diagnosis.