E1564. Role of Machine Learning in the Detection of Clinically Significant Prostate Cancer on Multiparametric MRI
  1. Solomon Kim; University of Rochester Medical Center
  2. Thomas Osinski; University of Rochester Medical Center
  3. Gary Hollenberg; University of Rochester Medical Center
  4. Eric Weinberg; University of Rochester Medical Center
The traditional approach to prostate cancer detection includes prostate specific antigen, digital rectal exam, followed by a transrectal ultrasound guided systematic biopsy. The use of multiparametric MRI (mpMRI) and MR/US fusion targeted biopsy have improved detection. However, mpMRI may miss some clinically significant prostate cancers, specifically those with cribriform pathology. The purpose of the study is to develop an artificial intelligence (AI) algorithm based on mpMRI to detect previously unrecognized clinically significant prostate cancers.

Materials and Methods:
Chart review and data collection was performed on approximately 360 IRB approved subjects who had undergone mpMRI and fusion biopsy at our institution. Inclusion criteria included: males older than 18 yo who had undergone mpMRI with fusion biopsy. Exclusion criteria included patients without available pathology. Each prostate specimen was re-examined by genitourinary pathologists and cancer lesions were plotted onto prostate maps. The prostate maps and mpMRI images were provided to Qmetric Technologies, who created a logistic model using local fractal dimension and wavelet decomposition signatures from T2 weighted images.

The average age of the patient in the demographic pool was 65. The average PSA preceding radical prostatectomy was 10.7 ng/ml. The average prostate volume was 44 cc. Prostate cancer lesions were found to have different contextual properties compared to benign regions on T2 weighted imaging using the AI algorithm. The area under the curve receiver operating characteristics was .77.

Machine learning is increasingly utilized in radiology and medicine as a whole. With regard to mpMRI and prostate cancer, utilization of machine learning can help us improve the sensitivity and specificity of mpMRI for detecting those lesions currently missed by radiologists. Improving the ability of mpMRI to accurately and consistently detect clinically significant prostate cancers with cribriform pattern 4 subtype could improve the accuracy of mpMRI in clinical practice, improve patient care and provide better therapeutic efficacy of focal therapies.