INTERNATIONAL CONGRESS OF
CancerGenetics
Artificial Intelligence for breast cancer detection
-
Fereshteh Arefi,1,*
1. Biology Department, Faculty of Biosciences, Tehran North Branch, Islamic Azad University, Tehran, Iran
- Introduction: AI technology is increasingly vital for breast cancer diagnosis due to rising patient numbers and limited pathologist resources. Computer-aided diagnosis (CAD) tools and deep learning (DL) methods have shown promise in enhancing accuracy and early detection. Studies like HASHI have demonstrated AI's effectiveness in handling large histopathology images and classifying malignancy levels. AI systems also assist in identifying metastasis locations and improving diagnostic efficiency. Despite challenges, integrating AI into clinical workflows shows potential for enhancing diagnostic accuracy and efficiency, ultimately aiding in early breast cancer detection and reducing avoidable deaths. For decades, research has focused on automated breast cancer detection in mammography to support radiologists. Despite advancements like digital mammography (DM) and digital breast tomosynthesis (DBT), radiologists' evaluations remain crucial. Even experienced radiologists can miss cancerous lesions or recall healthy women, adding to their workload. Since the 1990s, CAD systems using machine learning (ML) have helped radiologists by highlighting suspicious areas, but high false positive rates limited their acceptance. Recently, DL AI systems have shown significant improvements in detection accuracy, potentially enhancing or partially replacing radiologists' roles. This review highlights AI's development, benefits, and ongoing challenges in breast cancer detection.
- Methods: The evaluation of breast AI technology focuses on comparing AI techniques with traditional methods. Advances in computational power and digital data have significantly improved AI in medical imaging. AI includes machine learning (ML) and deep learning (DL), each with unique methods. Traditional computer-aided detection (CADe) and diagnosis (CADx) tools from the 1990s used ML to process image features, requiring extensive manual input. In contrast, DL techniques like deep neural networks (DNN) and convolutional neural networks (CNN) autonomously learn from data, improving tasks like image segmentation and classification with minimal human intervention. Training AI systems relies on high-quality data, involving thorough pre-processing and maintaining patient privacy. AI models use various learning paradigms, including supervised, unsupervised, and reinforcement learning. Effective training also involves data augmentation and transfer learning to overcome dataset limitations. Validating AI systems requires independent, generalizable models using diverse datasets to ensure fairness and robustness. Explainable AI (XAI) techniques, such as attribution maps, enhance transparency in AI decision-making, building trust and facilitating clinical integration by providing clear insights.
- Results: Deep learning (DL)-based AI systems have significantly improved breast cancer detection by increasing screening accuracy and reducing false positives and negatives. These systems can identify subtle abnormalities that human observers might miss. However, challenges like inconsistent datasets, biases, and regulatory issues hinder widespread adoption. Choosing the right AI system for clinical use requires evaluating factors such as accuracy, application, and compatibility with local populations and equipment. Regular audits are recommended to ensure safety and effectiveness, and access to performance metrics and training data is essential for informed decisions. Future AI advancements should incorporate diverse data sources, including historical and contralateral images, to enhance predictive accuracy. Techniques like federated learning and increased clinical data availability will support these improvements. Recent trials, such as MASAI and ScreenTrustCAD, have shown that AI can improve cancer detection rates and reduce recall rates. The ongoing AITIC trial suggests AI could automate the review of low-risk exams. The main challenge is deciding whether to use these results or conduct local trials to adapt AI for specific screening needs.
- Conclusion: AI technology shows great potential for improving breast cancer screening by enhancing accuracy and reducing radiologists' workload. Deep learning (DL)-based AI systems have outperformed traditional methods, sometimes even surpassing human radiologists. These systems help reduce observer variability and minimize false positives and negatives. To integrate AI effectively into clinical practice, standardized guidelines and reliable practices are necessary to ensure fairness and robustness. Ongoing research and validation are crucial to build clinical trust and confirm AI's efficacy. Collaboration among researchers, clinicians, and regulatory bodies is essential to address challenges and implement AI in screening programs. While traditional computer-aided detection (CAD) systems had high false positive rates, recent DL advancements have improved performance. High-quality data and effective preprocessing are vital for AI training, using various learning paradigms and data augmentation techniques. International initiatives are working on guidelines to enhance AI transparency and reliability. Continued research and validation are key for successfully integrating AI into routine breast cancer screening.
- Keywords: Artificial intelligence, Breast imaging, Breast cancer, Mammography, Screening
