Implementation Of Convolutional Neural Network(CNN) Resnet-50 Architecture For Mammography Image Classification In Early Breast Cancer Detection

Abstract

Breast cancer is one of the leading causes of death among women. Early detection through mammography screening is crucial to reduce mortality rates. However, manual image interpretation by radiologists is susceptible to subjectivity and visual fatigue, which can lead to misdiagnoses (False Positives and False Negatives). This study aims to design and implement a deep learning model using a Convolutional Neural Network (CNN) with the ResNet-50 architecture to classify mammography images into three tissue conditions: Benign, Malignant, and Normal. The model was developed using a Transfer Learning approach based on ImageNet weights, with specific modifications made to the classifier head layer. The secondary dataset was sourced from Kaggle, resulting in a final total of 23,939 images after undergoing preprocessing stages that included background removal (Auto-Crop), dimensional resizing (224x224 pixels), normalization, and geometric data augmentation. Model training was optimized using a Two-Stage Training strategy consisting of Feature Extraction and Fine-Tuning—to overcome potential performance degradation caused by an imbalanced data distribution (class imbalance) and to maximize the recognition of pathological features. Testing results conducted on 4,542 images in the testing data demonstrated that the proposed model achieved an Overall Accuracy of 98%. The evaluation of the F1-Score metric recorded a value of 0.98 for the Benign and Malignant classes, and a perfect 1.00 for the Normal class. Furthermore, the model exhibited excellent sensitivity (Recall) with zero False Negative predictions for malignant cases detected as Normal. Overall, this modified ResNet-50 architecture is proven to be reliable, precise, and holds significant potential to be implemented as a Computer-Aided Diagnosis (CAD) system to support the efficiency of medical professionals' clinical decisions.