Performance of a deep learning algorithm on deep learning reconstruction-accelerated prostate biparametric MRI: A preliminary study.

Purpose

To evaluate the robustness of a deep learning algorithm (DLA) for prostate lesion detection and classification trained on conventional biparametric MRI (bpMRI) to changes in imaging inputs introduced by deep learning reconstruction (DLR)-accelerated MRI.

Materials And Methods

This retrospective study included 141 men (median age 70.5 years; median prostate-specific antigen [PSA] level 6.6 ng/mL). A DLA for prostate cancer (PCa) detection and classification was evaluated using four bpMRI input configurations acquired within the same examination: conventional T2-weighted imaging (C-T2WI) plus conventional diffusion-weighted imaging (C-DWI; reference), C-T2WI + DLR-DWI, DLR-T2WI + C-DWI, and DLR-T2WI + DLR-DWI. A definitive diagnosis was established in 115 patients using histopathology or ≥2 years of PSA follow-up; clinically significant PCa was defined as a Gleason grade group ≥2. Performance was assessed at the patient and lesion levels using diagnostic metrics, DLA level of suspicion (LoS), false positives per case (FP/case), and intersection over union (IoU).

Results

Per-patient diagnostic performance estimates were generally close across the input configurations, although the magnitude of the differences varied across metrics. With C-T2WI + DLR-DWI, the number of detected lesions decreased compared with the reference configuration (99 vs. 131), accompanied by lowerFP/case (0.478 vs. 0.722; P = 0.003) and greater IoU (0.194 vs. 0.153; P = 0.014). Differences in lesion-level sensitivity and positive predictive value were relatively small across configurations.

Conclusion

The evaluated DLA showed generally close performance estimates across conventional and DLR-accelerated bpMRI inputs. However, the vendor-specific nature of both the DLA and DLR limits generalizability and warrants further evaluation.