Artificial intelligent-powered detection of breast cancer

Abstract

Breast cancer (BC) is characterized as uncontrollable growth of breast cells. Accurate screening of patients suspected with BC is crucial for timely treatment and minimizing cost. Medical expert relies on several techniques such as biopsy, ultrasound, mammography, Magnetic Resonance Imaging (MRI) etc. Despite reliance on these techniques, majority have drawbacks which include high cost, miss-diagnosis and miss-interpretation, false positive results. As result of the growing number of patients diagnosed with BC, radiologists are facing increase in workload which can delay diagnosis and increase susceptibility or prone to error. Integrating Artificial intelligence techniques into routine pathology practice have shown to reduce workload, errors and improve diagnostic efficiency. Therefore, in this study, we proposed the application of ensemble Deep Learning and Machine Learning approach for the detection of BC from both histopathological and ultrasound images. We curated 4 datasets from Kaggle repository which include 2 histopathological datasets (Breast Cancer Dataset (BCD) with 7783 images and BreaKHis 400X dataset (BH400X D) with 1693 histopathological images) and 2 ultrasound datasets (Ultrasound Breast Classification Dataset (UBCD) with 9016 images and Breast Ultrasound Images Dataset (BUID) with 1578 images. The acquired images are processed via resizing and colour enhancement and trained using customized CNN (D-ResNet) and 5 pre-trained models (ResNet101, ResNet50, VGG16, VGG19 and MobileNet) coupled with Random Forest (RF). Evaluation of the models based on unseen dataset resulted in 95.98 % accuracy using D-ResNet-RF validated on BCD, 92.34 % accuracy using D-ResNet-RF validated on BHX400, 84.82 % accuracy using ResNet101-RF validated on UBCD and 94.32 % accuracy using VGG19-RF validated on BUID.