Breast cancer remains one of the leading causes of cancer-related deaths among women worldwide. However, current mammography screening faces challenges such as limited accuracy, time-consuming diagnoses, and a shortage of radiologists. These issues are particularly pronounced in younger women with dense breast tissue, where false-negative results can lead to delayed diagnosis and treatment.
To address these problems, a collaborative team led by Dr. Po-Sheng Yang of Mackay Memorial Hospital, Distinguished Professor Jia-Ching Wang, and Professor Yi-Chiung Hsu of National Central University has developed an intelligent mammography diagnostic and automated reporting system. By leveraging advanced image processing, deep learning, and natural language processing (NLP) technologies, the system significantly enhances the efficiency of breast cancer detection. It reduces the traditional report generation time from 15–55 days to just 30 minutes while accurately identifying abnormal lesions, thereby improving both diagnostic efficiency and accuracy.
AI Enhances Early Breast Cancer Detection and Diagnostic Precision
This system integrates deep learning with multimodal analysis techniques, including image enhancement, image recognition, and NLP, to perform precise analysis of mammograms. It automatically detects and marks potential abnormal areas, assisting radiologists in making faster and more accurate interpretations. By utilizing a hybrid architecture that combines convolutional neural networks (CNN) with transformer models, the system classifies tumor edges and shapes, effectively distinguishing normal tissue from pathological regions. Additionally, the integration of a visual attention mechanism allows the system to pinpoint and map the location and distribution of calcifications, significantly improving diagnostic sensitivity.
The system can also be seamlessly integrated into existing medical infrastructures, such as hospital PACS systems and mobile breast screening units, enabling high-quality breast cancer screening services even in remote and underserved areas.
To further enhance diagnostic accuracy, the team employed semi-supervised learning strategies, combining labeled and unlabeled data to continually improve model performance. Image augmentation techniques were also applied to significantly enhance contrast, making subtle lesions more visible. Furthermore, the introduction of meta-learning networks supports few-shot learning on limited datasets, greatly improving classification accuracy. Empirical data demonstrates that the system achieves a detection accuracy of over 90% for breast masses and a benign-malignant classification accuracy of up to 75%.
Accelerating Industrialization and Advancing Breast Cancer Prevention and Healthcare Resource Optimization
Dr. Po-Sheng Yang noted that this diagnostic system not only shortens breast cancer screening and interpretation times and improves diagnostic accuracy but also expands screening coverage in rural areas, helping to address healthcare resource disparities. Looking ahead, the team plans to continue optimizing the system’s performance, accelerating its industrialization, and collaborating with healthcare institutions and medical equipment manufacturers to promote this technology and enhance the overall efficiency of healthcare systems.
Resource: 智慧診斷乳房X光精準辨識病灶 30分鐘生成報告
