Radiologists, Urologists, Computer ScientistsCurrent clinical protocol to evaluate men with prostate cancer underestimates the extent of prostate cancer in 30% of patients; thus, there is increased reliance on multiparametric prostate MRI (mpMRI) to more accurately determine patient’s risk level. Interpreting prostate mpMRI is challenging due to inherent heterogeneity of the prostate and large volume of complex visual information. We tested a machine learning algorithm which uses routine mpMRI to provide simplified likelihood predictions of tumor locations using regional distribution of features (including shape and distribution of intensity) corresponding to the region of interest (i.e. tumor). This retrospective study determined the algorithm’s ability to detect prostate cancer in new cases.This is a retrospective HIPAA-compliant, IRB approved study. For evaluation, the model was tested on 7 patients diagnosed with solitary prostate cancer. Multi-parametric MRI maps were pre-processed as follows: images were co-registered to a single reference standard using 6 landmark points manually placed on each slice where the prostate was visible. The affine registration was inferred from the pairs of corresponding points and used to map the images to a common image space. Each prostate was manually segmented using a region of interest established using an image processing software (Osirix). Similarly, whole mount lesions were manually traced within the software following the written annotation, and were used to create binary masks of presence of cancer. The learning was achieved by extracting local image patches [1] (size 21x21) at similar locations in T2, high b-value DWI and ADC images. Each set of coregistered multimodal patch was combined into a single vector using concatenation and used as input to the regression model. The output was set as the presence of tumor seen on whole mount histopathology after robotic prostatectomy. A Kernel Spectral Regression (KSR) [2] was used to model the nonlinear relation between T2, high b-value DWI and ADC images and location of prostate tumor on whole mount. We performed a leave-one-patient-out cross-validation to detect presence of prostate cancer on new images.7 patients with solitary prostate cancer satisfied the inclusion criteria and were included in this study. The mean age was 64.5 years and PSA was 9.7 ng/ml. Preoperative TRUS Gleason score (GS) was 7 in 3/7 (43%) patients; GS8-10 in 4/7(57%) patients. The average prostate volume was 27.2 cc (SD 9.5) and average tumor diameter on MRI was 1.7 cm (SD 0.7); Preoperative mpMRI showed single suspicion lesion with PI-RAD score of 4/5 in 4 (57%) patients; and 5/5 in 3 (43%) patients. 3 of 7 (43%) patients had stage T2 disease; 2/7 (28%) had T3a and 1/7 (14%) had T3b disease. Postoperative whole mount histopathology showed GS3+4 in 4/7 (57%); 4+3 in 2/7 (28%) and GS 4+5 in 1/7 (14.2%) patients. The performance of the model was evaluated under a leave-one-patient-out cross-validation. The average accuracy was 71 ± 8% in detecting the cancer region. The automatic detection of cancer in prostate from multiparametric MRI could play a major role in the routine clinical diagnostic. Although a larger cohort of patients should be studied, we have demonstrated, as a proof of concept, that automation using advanced machine learning could provide relevant insight about the likely location of the cancer tissue. Our preliminary results suggest that using a trained machine learning algorithm (spectral regression model) to analyze multiparametric is highly accurate for automatically localizing prostate cancer.