Breast cancer remains a leading health concern worldwide. Early detection is vital, yet challenges persist, especially when it comes to accuracy across different patient groups. Recently, Hologic announced that its Genius AI Detection 2.0 software performed equally well across Asian, Black, Caucasian, and Hispanic women in a study of over 7,500 patients. This is a significant stride toward equitable healthcare. But how does breast density affect AI accuracy? Let’s delve deeper.
The Impact of Breast Density on AI Accuracy
Artificial intelligence has shown great promise in detecting breast cancer early. However, breast density poses a significant hurdle. Dense breast tissue can mask tumors on mammograms, making detection harder for both radiologists and AI algorithms.
A study by experts from Duke University Medical Center analyzed nearly 400 digital breast tomosynthesis (DBT) exams using an FDA-approved AI algorithm. They found that the algorithm had an 11% false-negative rate. Alarmingly, 62% of these false negatives occurred in women with heterogeneously and extremely dense breasts. This pattern held true even after adjusting for age, race, and ethnicity.
Moreover, dense breast tissue doesn’t just challenge AI; it affects human interpretation too. Radiologists often struggle with dense breasts, leading to missed diagnoses. Therefore, training AI algorithms on diverse datasets that include various breast densities is crucial. Only then can these tools reliably assist in clinical practice.
Hologic’s Genius AI Detection 2.0 Performance
Hologic’s recent announcement brings encouraging news. In a study conducted by the Northwestern Feinberg School of Medicine, the Genius AI Detection 2.0 software demonstrated consistent performance across different racial and ethnic groups. The study evaluated 7,519 DBT examinations, including 2,532 cancer cases, from women identifying as Asian, Black, Hispanic, or white.
The AI’s sensitivity ranged from 88.6% in Black women to 91.5% in Hispanic women. Specificity varied from 57.6% in Hispanic women to 61.2% in Black women. Notably, the software performed best for Asian women, with 95.2% sensitivity and 71.6% specificity.
Sarah M. Friedewald, MD, Vice Chair for Women’s Imaging at Northwestern, emphasized the importance of this research. She noted that while AI solutions have been leveraged for years, understanding their performance across different groups is essential to minimize inherent biases.
The Importance of Diverse Datasets
The effectiveness of AI algorithms heavily depends on the data they’re trained on. If the datasets lack diversity, the AI may not perform well across all patient groups. This can exacerbate existing health disparities.
Breast density is a prime example of a variable that can influence AI accuracy. Since nearly half of all women have dense breasts, it’s imperative that algorithms are trained on images reflecting this reality. Furthermore, including diverse racial and ethnic groups ensures that the AI doesn’t inadvertently favor one group over another.
Erik Anderson, President of Breast and Skeletal Health Solutions at Hologic, highlighted the company’s commitment to equitable solutions. He pointed out that Black women in the U.S. have a 40% higher death rate from breast cancer compared to white women. By ensuring their technology performs effectively for all patients, Hologic aims to address such disparities.
Conclusion
Breast cancer detection is evolving with the advent of AI. Yet, challenges like breast density and the need for diverse training data remain. Hologic’s Genius AI Detection 2.0 software’s consistent performance across different populations is a promising development. However, continuous efforts are needed to train algorithms on varied datasets, ensuring accuracy for all women, regardless of race, ethnicity, or breast density.
Transitioning into a future where AI aids radiologists effectively requires collaboration and commitment. By focusing on inclusivity in data and acknowledging factors like breast density, we can move toward more accurate and equitable breast cancer detection.
