The Influence of Image Cropping Sizes on Mammographic Breast Cancer Classification Using CNN
Harshvardhan GM, Kei Mori, Sarika Verma, Lambros S. Athanasiou · 2023
Early-stage diagnosis of breast cancer can help plan a patient's prognosis more effectively, and also reduce the clinician's workload of manually inspecting mammograms, thereby augmenting their ability to diagnose. While various studies in the literature have proposed extracting regions of interest (ROI) and classifying them as “malignant” or “benign” using deep learning methods, seldom have investigated the impact of the ROI window size on the deep learning model's performance. In this paper, we investigate this aspect and propose the optimal ROI window size for training convolutional neural networks (CNN) on the Curated Breast Imaging Subset of the DDSM (CBIS-DDSM) database. We find that window sizes between$500\times 500$and$700\times 700$are most optimal for the ensemble CNN model used in our implementation.