Integrating Deep Learning Models and Data Augmentation Techniques for Improved Breast Cancer Detection
Mohammad R. HassanCommunications and Computer Engineering Department, Faculty of Engineering, Al-Ahliyya Amman University, Amman, Jordan. mhassan@ammanu.edu.jo0000-0002-2635-2181
Hamza Abu OwidaMedical Engineering Department, Faculty of Engineering, Al-Ahliyya Amman University, Amman, Jordan. h.abuowida@ammanu.edu.jo0000-0001-6943-6134
Qasem KharmaDepartment of Electrical Engineering, College of Engineering Technology, Al-Balqa Applied University, Amman, Jordan. ali.mohamad@bau.edu.jo0000-0001-9763-6029
Ali Mohd AliDepartment of Software Engineering, Faculty of Information Technology, Al-Ahliyya Amman University Amman, Jordan. q.kharma@ammanu.edu.jo0000-0003-4759-2835
Raheel AhmadSino-Pak Center for Artificial Intelligence (SPCAI): Institute of Applied Sciences and Technology (PAF-IAST), Mang, Pakistan. m21f0070ai011@fecid.paf-iast.edu.pk0000-0002-7189-2056
Mohammad AlhajComputer Tech Telecom, Canada. m.alhaj@ieee.org0000-0002-4517-8895
Keywords: Breast Cancer Detection, Histopathological Image Analysis, Deep Learning, Patch-based Classification, Multi-Magnification Analysis, Data Augmentation.
Abstract
Breast cancer (BC) is a prominent issue in global health that necessitates the utilization of progressively advanced diagnostic techniques to achieve early diagnosis and enhance patient outcomes. Patch-based histopathological images offer a deep insight into tissue structure and are crucial for accurately classifying benign and malignant breast tumours. Despite all the past work in this area, a helpful methodology is still lacking that connects different sources of data, amplifies the Input with various methods, and thoughtfully applies advanced deep-learning models to study these patches. We address BC detection here by working with various datasets, analyzing them in parallel at two magnifications (40X and 400X), and combining the results to explore better and diagnose the grouped tissues. A technique was employed to adapt data augmentation for the benign tumour class, addressing the issue of class imbalance. Every time, the DenseNet121 model predicted correctly at 40X, while the ResNet50V2 model was accurate when tested at 400X. When data at both magnifications is used in ResNet50, the model achieves a validation accuracy of 99.76%. They show that the best results are achieved when data from multiple sources and magnifications are consolidated. The model is made more general with augmentation, and pre-trained models play a part in detecting breast cancer (BC). Because it is more accurate, this style of testing enables patients to receive help when needed, thereby reducing the risk of unnecessary treatments.