Breast Cancer Characterization using Artificial Intelligence in Screening and Diagnostic Mammogram of Thai Women
Jaroonroj Wongnil, Anchali Krisanachinda, Rajalida Lipikorn · 2023
Breast cancer is a significant health issue in Thailand, necessitating accurate detection and characterization for effective treatment. This thesis focuses on leveraging artificial intelligence techniques to develop breast cancer screening and diagnosis using mammography images of Thai women. Furthermore, detecting and classifying mammography calcifications, crucial indicators of malignant or benign conditions had also been focused using RCNN and YOLOv4 in this study. AlexNet CNN was developed to increase the robustness of the RCNN model. AlexNet was tested with three different datasets, each consisting of 4,000 cropped ROI images. These datasets were divided into an 80% training set and a 20% validation set. 1,000 cropped ROI images were used as a test set. A five-fold cross-validation approach was implemented within the inner loop of the process across all datasets. The best performance AlexNet model was selected and used as the backbone for the RCNN model. The original YOLO4 was also used for studying. The ground truth of 5,000 bounding boxes from 3,265 mammogram images was created and consisted of malignant calcification in 2,500 bounding boxes and benign calcification in 2,500 bounding boxes. Divide dataset into 80% and 20% for the training and validation sets. The 1,000 bounding boxes were used for the testing set. The five-fold cross-validation was used with all datasets and performed. The parameters were fine-tuned in each round of the cross-validation. The learning process was completed when the model extracted specific features to represent malignant and benign calcification features for the tested mammography images. The research findings highlight the high performance of the models in distinguishing between malignant and benign calcifications. RCNN model 2 demonstrates remarkable precision and recall scores for malignant and benign calcifications. The precision for malignant calcifications is 0.82, and the recall is 0.84. For benign calcifications, the precision is 0.83, and the recall is 0.85. When considering both malignant and benign calcifications, the overall performance of the model, precision, recall, F1 score, and mAP, are 0.82(0.80-0.84), 0.85(0.83-0.87), 0.83(0.82-0.84), and 0.74(0.73-0.75) respectively. These findings demonstrate the efficacy of artificial intelligence in assisting radiologists with identifying and characterizing breast cancer in screening and diagnosis using mammogram images of Thai women. The high-performance models offer the promising potential to augment the workload of radiologists, particularly in regions with limited resources. Future studies can build upon this foundation to enhance the accuracy and efficacy of object detection models in the context of breast cancer characterization for the Thai women.