Diagnosis of Clinically Significant Prostate Cancer in Multiparametric MRI with Pseudo-Localization of Suspected Lesion.
Authors
Affiliations (3)
Affiliations (3)
- Shanghai University, No.99, Shangda Road, Shanghai, Shanghai, 200444, China.
- Huashan Hospital Fudan University, No,12 Wulumuqi Middle Road, Shanghai, Shanghai, 200040, China.
- Shanghai University, No.99,Shangda Road, Shanghai, Shanghai, 200444, China.
Abstract
To develop a two-stage diagnostic framework using pseudo-localization for patient-level diagnosis of clinically significant prostate cancer (csPCa) on multiparametric MRI (mpMRI), enabling better utilization of examinations without manual lesion annotations.
Approach. We included 1494 mpMRI examinations from the PI-CAI dataset, partitioned at the patient level into 217 annotated csPCa, 202 unannotated csPCa, and 1075 non-csPCa cases. First, a prostate transformer U-Net (PTUnet) was trained on the 217 annotated cases via five-fold cross-validation to generate pseudo-locations for the 202 unannotated csPCa cases. These pseudo-locations, along with expert lesion annotations and non-csPCa labels, were used to train an anisotropic UX-Net (AUX-Net). The Probability-Ushered Lesion Selection (PULSE) procedure derived patient-level diagnostic probabilities. We evaluated PTUnet, AUX-Net, PULSE, and the pseudo-localization strategy against representative models and baselines, and analyzed the link between pseudo-localization quality and diagnostic performance.
Main results. PTUnet achieved a Dice similarity coefficient (DSC) of 64.96±4.22% and an average precision of 65.69±5.45%, the best overall pseudo-localization performance among evaluated models. All four diagnostic models showed higher AUCs with pseudo-locations. AUX-Net with PULSE achieved an AUC of 83.11%, sensitivity of 82.35%, specificity of 75.75%, and Youden's index of 58.10%, compared with the 80.96% AUC of AUX-Net without pseudo-locations. Pseudo-locations from nine different models all improved diagnostic AUC over the no-pseudo-location baseline, indicating the benefit is not model-specific. Furthermore, localization DSC was significantly and positively correlated with diagnostic AUC (Spearman's ρ = 0.800, p = 0.0096), suggesting higher-quality pseudo-localization leads to better diagnosis.
Significance. The framework allows csPCa examinations without manual annotations to contribute to diagnostic model training. Results show pseudo-localization improves patient-level diagnosis across different architectures, and its quality is positively associated with downstream performance. The framework shows potential for non-invasive AI-assisted csPCa diagnosis and biopsy triage, pending further validation on independent larger-scale datasets and prospective clinical cohorts.