MP46-02: Detailed Analysis of MRI Concordance with Prostatectomy Histopathology Using Deep Learning-Based Digital Pathology

Sunday, September 12, 2021 8:00 PM to 10:00 PM


Authors: Lukas Hockman, Richard Fan, Bogdana Schmidt, Indrani Bhattacharya, Mirabela Rusu, Geoffrey Sonn

Introduction: Focal therapy for prostate cancer relies heavily on prostate MRI, highlighting the weaknesses of MRI alone for cancer detection. Prior studies comparing MRI to pathology relied upon tedious annotation and grading of cancer foci. Application of AI algorithms to digital pathology images enables rapid identification and grading of all cancer foci. We used a deep learning algorithm to perform individual gland level grading of digital whole mount prostatectomy slides, and register the pathology to preoperative MRI in 3D. Using this dataset, we sought to rigorously measure concordance in tumor detection between preoperative MRI and postoperative histopathology.

Methods: We analyzed MRI and histopathology findings from 160 axial sections in 30 men after prostatectomy. Whole mount prostatectomy specimens were axially sectioned using customized 3D printed molds and registered to MRI images using a state-of-the-art 3D approach (Figure 1). Cancers were identified and graded using a validated deep learning algorithm (DeepDx, Deep Bio, Seoul, South Korea). Cancerous tissue was classified as either aggressive (Gleason pattern 4-5) or indolent (Gleason pattern 3).

Results: We identified 88 distinct cancers in the 30 men. Of those, 30 (34%) were prospectively identified on preoperative MRI. Most prostates had multiple tumor foci; only 8 (27%) men had a single tumor. In MRI tumors, median tumor diameter was 20mm and mean proportion of aggressive cancer was 33%. In the 58 (66%) cancers missed on MRI, median tumor diameter was 5.5mm and mean proportion of aggressive cancer was 9%. Most of the missed cancers were small and low grade; 85% were <10mm and 75% contained <10% Gleason 4-5 disease. On per-patient analysis, the dominant tumor was detected in 26 (87%) cases and the highest-grade cancer in 25 (83%) cases.

Conclusions: Merging gland level grading by deep learning with 3D registered MRI findings enables a quantification of focal prostate cancers that is not possible when relying on human pathologists. MRI identified the index lesion in ~85% of cases; missed lesions were small or low grade. Based on these findings, men considering focal therapy for prostate cancer should be advised that most will continue to harbor small, nonaggressive cancers after treatment and will require surveillance.

Source of Funding: Departmental

Therapeutic Area
Oncology: Prostate