Deep Learning-Based Recommendation System for Breast Cancer Diagnosis

Breast cancer is considered one of the deadliest cancers among females. Despite the advanced achievement in the field of medical imaging analysis, a few early research works proposed semi-automatic machine learning algorithms that were complex and computationally expensive. Recently, developing a system based on deep learning concepts was the center of the attention to analyze mammograms, which is the golden standard imaging technique to diagnose the existence of an abnormality in breast tissues. Systems based on deep learning are still considered to be limited due to insufficient datasets. In this study, numerous experiments were made on small size ROI samples of mammogram images in the CBIS-DDSM dataset to find the best configuration of a pre-trained Convolutional Neural Network with ImageNet dataset. The main training concept is about extracting standard features automatically by striding filters over the input matrix (mammogram). Thus, a larger number of inputs lead to recognize a useful pattern to classify the abnormality. The pre-trained models along with data augmentation algorithms are applied to minimize the dataset challenge in the breast cancer diagnosing field. On the other hand, image preprocessing techniques helped to enhance the image inputs. The study addresses two pre-trained models VGG16 and ResNet50, both were fine-tuned in different depths and hyper-parameters. The experimental results obtained with good performance used a VGG16 pre-trained model after fine-tuning the last fully connected layer. The proposed VGG16 model outperformed other deep learning algorithms with an F1 score and AUC of 82% when classifying the abnormality type into calcification/ mass. The model has also a high score with a mean AUC equal to 0.80 when classifying the mammograms into four classes: benign calcification, malignant calcification, benign mass, and malignant mass. The final application in this study tries to assist radiologists to accomplish more precise decision on the abnormality pathology of breast lesions present in full mammogram images. The application also helps in reducing diagnosing time hence increasing the early detection time. Keywords: CNN, Transfer Learning, Breast Abnormality, Classification, Mammograms, Calcification, Mass.