INTRODUCTION

Convolutional neural networks (CNN) have been established as a powerful computational approach in computer vision tasks [1]. The U-net architecture implemented as an encoder-decoder with skip connections can maintain image resolution while providing segmentations into tissue classes [2]. Automated detection and segmentation of suspicious lesions on MRI carries high potential for clinical diagnosis of prostate cancer (PCa) [3]. Here, we train a diffusion-weighted imaging (DWI) only U-net and compare its performance with full multiparametric clinical PI-RADS assessment.

METHODS

Significant prostate cancer (sPC) was defined as Gleason Grade Group >=2 PCa. Clinical MR-lesions targeted by fusion biopsy in 316 consecutive patients examined on a single 3T MRI system were manually segmented. A U-net was trained with segmentations of confirmed sPC on biopsy. 63 patients were held out as an independent test set. Training proceeded on the remaining patients in a 4-fold 15 times repeated cross-validation using a 190/63 patient-ratio for training/validation in each fold. Individual segmentations of the resulting CNN ensemble (CNNE) contributed equally to the final voxel-wise tumor probability score (TPS= number of nets including voxel in tumor segmentation / total number of nets). At lesion-based analysis, an overlap of manual and CNN segmentation of at least one voxel was considered a match. Ground truth for patient-wise performance assessment was sPC presence on a combined targeted and extended 24-core systematic TRUS/MRI fusion biopsy map. Receiver operating characteristics curves were generated using available TPS and PI-RADS values. TPS thresholds were adjusted to match radiologist sensitivity at PI-RADS >=3 and >=4. The final ensemble of 60 cross-validated networks was applied to the test set and the performance of the ensemble was evaluated against the radiologist at the previously selected thresholds.

RESULTS

On a patient basis in the validation set of 253 patients, radiologists at PI-RADS>=3 and CNNE at TPS>=0.13 achieved a sensitivity of 0.94 (88/94), with specificity of 0.22 (35/159) for radiologists and 0.15 (23/159) for CNNE. At TPS>=0.6 CNNE reached a sensitivity of 0.85 (80/94) and a specificity of 0.45 (72/159) while radiologists at PI-RADS>=4 had a sensitivity of 0.83 (78/94) and a specificity of 0.53 (84/159). On a patient basis in the test set of 63 patients, radiologists at PI-RADS>=3 and CNNE at TPS>=0.13 achieved a sensitivity of 0.96 (26/27), with specificity of 0.22 (8/36) for radiologists and 0.14 (5/36) for CNNE. At TPS>=0.6 CNNE reached a sensitivity of 0.82 (22/27) and a specificity of 0.44 (16/36) while radiologists at PI-RADS>=4 had a sensitivity of 0.85 (23/27) and a specificity of 0.50 (18/36). Radiologists at PI-RADS>=3 identified 88 lesions in 54/63 test patients and 58 lesions for PI-RADS>=4 in 41/63 patients, while CNNE identified 205 lesions in 58/63 test patients for TPS>=0.13 and 91 lesions in 42/63 patients for TPS>=0.6. 33 radiological lesions were sPC positive on targeted biopsy. Of these, 21 (64%) were identified by CNNE at TPS>=0.13 and 18 (55%) at TPS>=0.6. Of the remaining 55 sPC-negative radiological lesions, 27 (49%) were indicated at TPS>=0.13 and 12 (22%) at TPS>=0.6. 49/88 (56%) of the total radiologists PI-RADS>=3 lesions were found by CNNE at TPS>=0.13, while 156/205 (76%) lesions were outside the regions identified by radiologists. 3/63 (5%) test patients had either negative targeted histopathology or no MR lesions, but positive total systematic histopathology (Gleason group 2). CNNE identified 6 additional lesions at TPS>= 0.13 in 2/3 (67%) and 1 additional lesion at TPS>=0.6 in 1/3 (33%) of these patients. For patients without radiologically identified lesions and negative histopathology for sPC, CNNE found additional lesions in 7/8 (88%) patients at TPS>=0.13 in 4/8 (50%) patients at TPS>=0.6. In 1/5 (20%) patients in whom CNNE at TPS>= 0.13 had found no lesions and 5/21 (24%) for TPS>=0.6, sPC was histopathologically confirmed (Gleason Grade 2).

DISCUSSION

CNNE performance at selected thresholds was stable in the test set compared to the validation and achieved a patient-based sensitivity close to clinical radiologist performance using the full multiparametric data, at slightly lower specifiticy.

CONCLUSION

A U-net using diffusion MRI information only for prostate lesion segmentation and detection demonstrated a promising performance compared to clinical radiologist interpretation, suggesting its potential as a triage tool. These findings highlight the ability of computer vision to access and use salient information from prostate MRI and encourage development of the method in larger cohorts.

Acknowledgements

No acknowledgement found.