AI-Enhanced Mammograms Boost Cancer Detection by 23%

- The AI technology, called ProFound AI and developed by iCAD, utilizes deep-learning convolutional neural networks. It was trained on a massive dataset of over 20 million images from more than 130 facilities, including 13,500 biopsy-proven cancers, to recognize and flag suspicious soft tissue densities and calcifications. - ProFound AI functions as a "second read" for radiologists. The system analyzes each mammogram image and assigns a "Certainty of Finding" score to suspicious lesions and an overall "Case Score" indicating the likelihood of malignancy, helping to prioritize cases. - In a direct comparison with another AI system, Transpara, ProFound AI demonstrated significantly better performance in breast cancer detection, with a higher Area Under the Curve (AUC) of 0.93 versus 0.86. However, studies show that a standard double reading by two human radiologists still outperforms both AI systems. - A key benefit for healthcare providers is a dramatic reduction in the time radiologists spend reading scans. Studies have shown that reading time can be cut by more than half, from an average of 64 seconds to just 30.4 seconds per case when using the AI assistance. - The system received its initial FDA clearance in December 2018, becoming the first AI software for 3D mammography (tomosynthesis) to do so. Most AI-enabled medical devices are cleared through the FDA's 510(k) pathway, which requires demonstrating "substantial equivalence" to a legally marketed device that is not subject to premarket approval. - This AI-enhancement is being offered as an optional service to patients at Providence St. Joseph Hospital for an out-of-pocket fee, which is a common model as insurance reimbursement for AI analysis is not yet standard. - For developers and entrepreneurs in the AI/ML space, building and launching a medical AI product involves significant regulatory hurdles. An indie hacker who built a medical imaging AI business noted that the FDA Class II approval process, following ISO and IEC standards, took their company nine months and cost under $70,000 using a consultant. - While effective, the technology is not foolproof and is vulnerable to adversarial attacks. Researchers have demonstrated that it's possible to use generative adversarial networks (GANs) to manipulate medical scans, adding or removing cancerous findings in a way that can fool both the AI and human radiologists.