AUA 2017: Quantitative Digital Image Analysis and Machine Learning Accurately Classifies Primary Prostate Tumors of Bone Metastatic Disease Based on Histomorphometric Features in Diagnostic Prostate Needle Biopsies
They focused on PC patients with high-risk disease. Indeed, 15% of men who present with PC harbor high-risk disease, and the 5yr-survival for M1 patients is dismal (28%) compared to M0 patients (100%). Current radiologic technology has virtually no ability to detect micrometastatic disease, which means some men with micrometastatic disease undergo curative-intent therapy with possibly no benefit.
Eric Miller from UCLA presented a pilot study using quantitative digital imaging (QI) to develop algorithms that can help classify M0 vs. M1 disease based on histopathologic characteristics of prostate needle biopsies obtained at the time of PC diagnosis. They retrospectively analyzed 40 M0 patients and 69 M1 patients whose diagnostic biopsy slides were available for digital analysis.
The H&E slides were digitally scanned at 40x and images were processed in a manner that 500 nuclei per case were available for analysis. Algorithms were used to extract nuclear features of the PC cells, sort the features to categorize the variations, place values on those features, then rank the feature values. The training images were analyzed by machine learning (artificial intelligence) algorithms to find the strongest features capable of distinguishing M0 from M1 patients. Incredibly, the derived algorithms had up to a 96% discriminatory ability to classify patients correctly! This was then validated on a separate set of biopsy images and the strength of the algorithms were confirmed.
This is a highly multidisciplinary approach to resolving a complex question previously out of reach for clinicians. Combining the knowledge gleaned from pathology, radiology, and informatics and incorporating all of this information into machine learning algorithms is, frankly, the biggest leap forward in medicine since the development of a microscope. This research shows that complex biological information can be gleaned from something as simple as an H&E side. Information previously invisible to the human eye is quickly apparent to the artificial intelligence eye. I congratulate this multidisciplinary team for helping to keep Urology at the forefront of this field.
Presented By: Eric Miller; UCLA, Los Angeles, CA
Written By: Shreyas Joshi, MD, Fox Chase Cancer Center, Philadelphia, PA
Twitter: @ssjoshimd
at the 2017 AUA Annual Meeting - May 12 - 16, 2017 – Boston, Massachusetts, USA