From Scan to Treatment Plan, AI Helps Close Breast Cancer’s Deadliest Gaps

Breast cancer is the most commonly diagnosed cancer among American women, but the system around it is slow and uneven. Many women over 40 still skip yearly screening. Radiologists are handling more mammograms with fewer colleagues. And when cancer is found, the tests that shape treatment can take weeks to come back.

NVIDIA’s Inception startups are aiming at those bottlenecks from several angles. On the imaging side, iSono Health has built ATUSA, an FDA-cleared wearable ultrasound system that captures a standardized 3D breast volume in about two minutes per breast. A conventional handheld ultrasound can take up to 45 minutes, and the company says its AI-assisted scan is 28% more sensitive than handheld 2D ultrasound. It was trained on thousands of full-breast scans and more than 1.5 million ultrasound frames, and it’s already available through partner clinics in California, Texas, Georgia, Tennessee and Washington D.C.

That consistency matters as much as speed. Handheld ultrasound depends on who is holding the probe, which makes year-to-year comparisons messy. ATUSA is designed to scan the whole breast the same way every time, reducing operator variability and creating a repeatable record clinicians can use to track change. iSono is also building out lesion detection, 3D segmentation and lesion classification, with a multicenter study of 3,200 patients underway at sites including UC Davis and Vanderbilt University Medical Center.

Whiterabbit.ai is tackling the mammogram side of the problem. Its FDA-cleared WRDensity software automatically assesses breast density and has already been used in the care of hundreds of thousands of patients. The company also has WRRisk for long-term breast cancer risk, and it’s working on new mammography AI aimed at catching more cancers while automatically clearing negative studies. That is the sort of dull, practical automation healthcare actually needs: less busywork for radiologists, fewer avoidable callbacks, and faster answers for patients. Its models are trained on NVIDIA GPUs at Washington University in St. Louis, with extra cloud capacity, and inference runs on NVIDIA GPUs in the clinic.

The last step is treatment, where waiting can be brutal. Ataraxis AI is using digital pathology slides and clinical data to predict how a cancer will respond to therapy, including whether presurgical chemotherapy is likely to shrink a tumor and a patient’s five-year recurrence risk after surgery. SimBioSys is doing something similar from another angle, building 3D models of tumors and soft tissue to guide surgery and treatment plans. Both are examples of the same basic idea: AI is most useful here when it shortens the gap between a scan, a diagnosis and a decision.