Early Detection of Breast Cancer from Mammogram Images Using Peafowl-Optimized Channel-Attention InceptionResNet Model

Abstract

Breast cancer is one of the most common and deadliest diseases affecting women globally. This paper proposes an enhanced Deep Learning (DL) system for detecting breast cancer through mammographic images. In the preprocessing phase, a bilateral filter is used to clean up background noise from the images while preserving critical anatomical and edge information of the breast. Regions-of-Interest (ROIs) are accurately segmented using a multi-scale K-means clustering method. Central architecture of this system is Channel Attention InceptionResNet, which functions as both a feature extractor and classifier. This architecture incorporates both residual connections and Inception modules to leverage their respective advantages. In order to enhance model's ability to focus on the most discriminative features, a Channel Attention Module is added using the Squeeze-and-Excitation (SE) mechanism to system. SE channel attention module applies a sequence of global average pooling operations, Multi-Layer Perceptron (MLP) processing, and sigmoid activations, enabling the model to adaptively reweight the importance of feature channels. After reducing redundant spatial dimensions by pooling the collected deep features with Global Average Pooling (GAP), dropout regularization is utilized to reduce the effects of overfitting. The last part is done by connecting all the features together inside a dense layer. This layer uses the ReLU function to analyse each feature to activate on an output level based on the observed outcomes. Then, final output layer uses a Softmax function to predict probability that a patient is in one of diagnostic categories (benign or malignant). Hyper-parameter tuning is achieved through the use of the Peafowl Algorithm (POA) to further enhance the accuracy of the model. To validate accuracy of developed model, comprehensive test experiments are conducted using Python examining performance on benchmark datasets for mammography. Compared with traditional techniques, the proposed framework proved to outperform by producing results with significantly higher accuracy of 0.9607, precision of 0.9810, Recall of 0.9599 and FI-Score of 0.9704.