RSNA 2016

Abstract Archives of the RSNA, 2016


IN229-SD-TUA6

An Investigational Patch-based Convolutional Neural Network Model for the Detection of Clinically Significant Prostate Cancer using Multiparametric MRI

Tuesday, Nov. 29 12:15PM - 12:45PM Room: IN Community, Learning Center Station #6



Awards
Trainee Research Prize - Medical Student

Karthik V. Sarma, BSc, Los Angeles, CA (Presenter) Nothing to Disclose
Xinran Zhong, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
King Chung Ho, MSc, los angeles, CA (Abstract Co-Author) Nothing to Disclose
Daniel J. Margolis, MD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Steven S. Raman, MD, Santa Monica, CA (Abstract Co-Author) Nothing to Disclose
Fabien Scalzo, PhD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Kyunghyun Sung, PhD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Nelly Tan, MD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Corey W. Arnold, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
PURPOSE

Despite prostate cancer being the second leading cause of cancer death in American men, the USPSTF recommends against screening to avoid overdiagnosis and treatment of indolent disease. The use of multiparametric magnetic resonance imaging (mp-MRI) has shown potential to discriminate aggressive from indolent disease. We demonstrate a convolutional neural network (CNN) that can generate a voxel-wise cancer probability map for clinically significant (Gleason score >= 7) prostate cancer.

METHOD AND MATERIALS

mp-MRI data was collected retrospectively for a set of 22 patients who had undergone radical prostatectomy. Surface mesh annotations were manually created from imaging for prostate and lesion segmentation. The prostate and tumors within were segmented by a genitourinary pathologist on whole mount histopathologic slides and manually mapped to the meshes. Ground truth was generated with any voxel contained within a lesion mesh with an assigned score >= 7 assigned to the positive class and all other prostate voxels assigned to the negative class.A CNN was trained on the resulting dataset with manually registered input channels for T2-weighted MRI, ADC, Kep, and Ktrans maps. 21x21 training patches were created for every prostate voxel with ground truth assigned based on the middle voxel. A four-layer model with two convolutional layers with max-pooling and two linear layers with dropout was used.

RESULTS

Leave-one-patient-out (i.e. 22-fold) cross-validation was performed, with results averaged across each fold. A softmax classifier was used for assignment to the lesion or normal class using a threshold calculated to give 80% specificity. For evaluation, voxels with a positive probability above the threshold value were assigned to the positive class. Average results were as follows: Accuracy 80%, AUC 0.72, S90 40%.

CONCLUSION

The patch-based CNN system was able to generate results competitive with voxel-based predictive systems. Future work includes training models with larger datasets, which could allow the use of deeper networks that may improve performance. Future models could also attempt to predict Gleason score directly instead of a dichotomized indicator and correct for motion and distortion.

CLINICAL RELEVANCE/APPLICATION

This investigation into the use of CNNs for detecting clinically significant prostate cancer in mp-MRI demonstrates that deep learning may be useful for prostate CAD.