ENHANCED BREAST CANCER DETECTION USING TRANSFER LEARNING: A COMPARATIVE STUDY OF VGG19, RESNET50, XCEPTION, AND INTEGRATED CLASSIFIER APPROACHES

Abstract

Breast cancer is distinguished by the abnormal and uncontrolled proliferation of cells within breast tissues. Certain irregularities may be missed or improperly detected because of the complicated patterns of breast abnormalities and the limitations of human visual judgment. In recent years, computer vision has developed as an important tool in healthcare, enabling applications such as disease diagnosis, tumor detection, medical imaging, and patient monitoring. While Convolutional Neural Networks (CNNs) have shown effective in image processing applications, their success is often dependent on vast amounts of training data. However, a lack of labeled medical imaging data makes it difficult to train CNNs from scratch for clinical applications. To overcome this, transfer learning is used, which allows pre-trained models to be adapted for effective medical picture categorization on little datasets.

This study uses three well-known pre-trained CNN models to categorize mammographic pictures as normal or abnormal: VGG19, Xception, and ResNet50. The collected features from these models were identified using two distinct machine learning techniques: Logistic Regression and a custom-designed Neural Network classifier. Each model's performance was measured using common measures such as , F1-score, accuracy, recall, precision, and confusion matrix analysis. Among the examined models, VGG19 in conjunction with the neural network classifier produced the greatest classification accuracy of 94%, surpassing ResNet50 and Xception, which had accuracies of 90% and 88%, respectively. VGG19's outstanding performance across various evaluation measures demonstrates its applicability for breast cancer diagnosis in the analyzed dataset.