High-level Features in Deeper Deep Learning Layers for Breast Cancer Classification

• Published in: Proceedings - 2021 11th IEEE International Conference on Control System, Computing and Engineering, ICCSCE 2021
• Main Author: Razali N.F.; Isa I.S.; Sulaiman S.N.; Karim N.K.A.; Osman M.K.
• Format: Conference paper
• Language: English
• Published: Institute of Electrical and Electronics Engineers Inc. 2021
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85116236071&doi=10.1109%2fICCSCE52189.2021.9530911&partnerID=40&md5=f974c267af6dfcd673a7684ceb0ce7d3

Early detection of breast cancer is crucial when treating than cure in later mammogram screening processes. To date, researchers extensively proposed the implementation of artificial intelligence to develop a computer-aided system (CAD) to determine types of breast tumour lesion, whether benign or malignant. This approach is significant to minimise the rate of misinterpretation in false positive and false negative diagnosis results among radiologists. Lack of established medical datasets publicly available has become the main reason why the system is not fully implemented in clinical settings yet. This study is aimed to investigate the performance of a convolutional neural network (CNN) to detect cancerous lesion types. The pre-trained CNN networks are tested on two established public datasets, CBIS-DDSM and INbreast. Pre-processing using denoising and contrast limited adaptive histogram equalisation (CLAHE) and augmented to lessen the effect of overfitting. The pre-trained CNNs AlexNet and InceptionV3 represent shallow and deeper neural networks respectively, trained using the transfer learning method. Performance of the system is tested and its accuracy, losses, sensitivity, specificity, and receiver operating characteristic curve (ROC) are evaluated. The InceptionV3 network performs better with the highest testing and area under the curve (AUC) at 99.93% compared to shallower AlexNet at 98.92% using INbreast dataset. Training the system using augmented data is proven to improve testing accuracy at 86.7% from 60.26% using a non-augmented dataset in low-quality input images. Meanwhile, using a shallower network for transfer learning produces high accuracy results without compromising computational cost. This study serves as the platform to improve the system's performance by varying the pretrained networks used and getting different features from each convolutional layer to be trained in the future. © 2021 IEEE.