Prostate biopsies are usually performed under ultrasound guidance and involve collecting 12 to 16 samples. However, these biopsy samples provide only limited glimpses into a much larger and more complex landscape. The scattered samples provide limited insight, leading to a miss rate of up to 52 percent of clinically significant cancers.

Effective prostate cancer treatment relies on early detection of cancer before it has spread outside the prostate. To help improve the success rate of finding prostate cancers during routine biopsy, Mirabela Rusu, PhD, Assistant Professor, Stanford Medicine Department of Radiology (IBIIS division), developed an artificial intelligence tool called ProCUSNet that analyzes the ultrasound images already acquired during the biopsy procedure.

The ProCUSNet model was trained on over 2,200 prostate cancer patients, using 3D volumes reconstructed from 2D B-mode images captured during routine transrectal ultrasound procedures. Unlike previous research that required raw or investigational imaging, this approach focused on real-world data already available in clinical settings.

The current version takes approximately 10 seconds to analyze a 3D ultrasound volume and was developed using data from a specific ultrasound system. Broader validation across different devices and practice settings will be necessary. Future studies will also explore faster integration, prospective clinical trials, and adaptation to other ultrasound modalities.

Dr. Rusu said “What makes ProCUSNet unique is its ability to localize areas of cancer on standard ultrasound images using deep learning. We’re not asking clinicians to change how they perform biopsies. Instead, we’re enhancing the value of the data they already collect with a tool that’s fast and practical for use in the real world.”

The findings, recently published in European Urology Oncology, show that ProCUSNet can help clinicians detect high-grade prostate cancers. The algorithm was able to detect 82% of clinically significant cancers and identify 44% more lesions than human readers interpreting the same ultrasound images.

The researchers also showed that ProCUSNet was also able to detect cancers missed by standard systematic biopsy. Among patients who went on to have surgery, nearly 30% had high-grade tumors that were not captured through conventional sampling. ProCUSNet successfully flagged many of these otherwise undetected lesions.

The potential clinical impact of this approach is substantial. As the majority of prostate biopsies worldwide continue to rely solely on conventional ultrasound, tools such as ProCUSNet may offer a practical means to improve diagnostic precision and consistency.