Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Prostate Cancer Malignancy Detection and Localization From MpMRI Using Auto-Deep Learning: One Step Closer to Clinical Utilization

W Zong1, E Carver2*, S Zhu3, E Schaff4, D Chapman5, J Lee6, I Chetty7, W Wen8, (1) Henry Ford Health System, Troy, MI, (2) Wayne State University, Troy, MI, (3) ,,,(4) Henry Ford Health System, ,,(5) Henry Ford Health System, ,,(6) Trinity Health, ,,(7) Henry Ford Health System, Detroit, MI, (8) Henry Ford Health System,

Presentations

PO-GePV-M-57 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Automatic diagnosis of malignant prostate cancer patients from mpMRI has been studied heavily. Model interpretation and domain drift have been the main roadblocks for clinical utilization. This abstract serves as one step closer towards clinical utilization where the trained model can be generalized to data from a different cohort and intensive delineation from medical experts has been greatly eased.

Methods: Data used in the study originated from the SPIE-AAPM-NCI Prostate Gleason Grade Group Challenge. Two modalities including T2 weighted, and ADC maps were used in the experiments. The training cohort consisted of patients where ADC maps were computed on DWIs with three b-values (50, 400, and 800 s/mm2). The testing cohort consisted of 40 patients where ADC was computed from two b-values (0 and 1,000 s/mm2). (2) The second source of domain drift came from the lack of contour of sub-regions in the testing cohort, including peripheral-zone (PZ) and central-gland (CG). (3) Also, number of slices were varied in the testing cohort. The solution we proposed for each of the challenge was (1) to train separately on the sub-regions, namely PZ-detector and CG-detector for different rendering of lesions on the two sub-regions. (2) to test both detectors on the same patient from the testing cohort. (3) to use two consecutive slices as the input and connect into the sequence for the patient.

Results: By making prediction for each slice and connecting the slices into the sequence, the PZ-detector and CG-detector were able to work together to highlight the suspicious slice out of the sequence. Grad-CAM was used to help locate the lesion within the prostate and verify the prediction accuracy.

Conclusion: Our strategy of training PZ-detector and CG-detector separately were verified robust to the domain drift issues caused by ADC map calculated from DWIs with different b values.

Keywords

Not Applicable / None Entered.

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email

Share: