Region of Interest-Based Breast Cancer Detection with Oversampling Technique
Defri Kurniawan, Abu Salam, Yani Parti Astuti, Catur Supriyanto, Guruh Fajar Shidik, Pulung Nurtantio Andono, Noor Zuraidin Mohd Safar · Ingénierie des systèmes d information · 2025
Breast cancer detection using medical imaging remains a challenging task due to the large volume of mammograms and the inherent class imbalance in datasets.This study proposes a novel regions of interest (ROIs)-based approach using RSNA screening mammography breast cancer detection dataset.By focusing on specific ROIs within the mammograms, the computational load is reduced while allowing the model to concentrate on the most critical areas.Additionally, SMOTE Tomek Link is applied to mitigate the class imbalance by generating synthetic samples for the minority (cancerous) class and removing noisy or overlapping samples.Three dataset splits were created: Split 1 (5:1 ratio of normal to cancer cases), Split 2 (3:1), and a fully balanced Random Under-Sampling (RUS) dataset.Various CNN models, including InceptionV3, ResNet152V2, DenseNet201, and EfficientNetB7, were evaluated on different dataset splits.Our results demonstrate that the EfficientNetB7 model, in conjunction with ROI extraction and SMOTE Tomek Link, achieves the highest accuracy of 97.41% on the Split 2 dataset, highlighting the effectiveness of these preprocessing techniques in enhancing deep learning-based breast cancer detection.