A multi-architecture study of specificity refinement and false-positive mechanism analysis in prostate MRI

Technical Summary: Multi-Architecture Study of Specificity Refinement and False-Positive Mechanism Analysis in Prostate MRI

Problem Statement
Clinical translation of prostate MRI-based artificial intelligence (AI) systems is currently hindered by high false-positive (FP) rates. While existing pipelines are heavily optimized to maximize sensitivity for clinically significant prostate cancer (csPCa) detection, specificity receives less attention. Consequently, AI-flagged regions often do not correspond to biopsy-confirmed cancer, leading to unnecessary follow-up imaging and biopsies. Two critical gaps exist in the current literature:

1. Methodological: Improving case-level specificity typically requires full model retraining. Lightweight, post-hoc recalibration methods (e.g., temperature or Platt scaling) are limited because they preserve ranking and cannot alter specificity beyond what simple threshold tuning achieves.
2. Mechanistic: The nature of residual false positives is rarely characterized. It remains unclear whether FPs are arbitrary model artifacts or if they share specific imaging properties with true cancers, which would fundamentally change how radiologists should adjudicate AI outputs.

Methodology
This retrospective study utilized two public cohorts: the PI-CAI dataset (1,500 multi-vendor studies, with a 5-fold cross-validation split) and the external Prostate158 dataset (158 studies). The research employed a frozen, multi-modal csPCa detection backbone (MIGF-nnUNet A2 variant) and introduced two primary analytical components:

1. Post-hoc Specificity Refinement (P2a): A lightweight, 89,216-parameter refinement head was trained on the frozen backbone's decoder features. This head operates as a residual module, adding a learned correction to the backbone's logits before the softmax layer. It was trained using focal loss and a mean-squared regularizer on PI-CAI fold-0, with hyperparameters fixed to ensure sensitivity preservation.
2. Evidence-Grounded FP Mechanism Analysis (P2b): To understand the nature of FPs, the study extracted three Region of Interest (ROI) types per case: ground-truth lesions, backbone-predicted FP regions, and contralateral benign ROIs. Raw imaging contrast ratios (T2-weighted, Apparent Diffusion Coefficient [ADC], and high-b-value) were computed for these regions against peri-lesional rings.
   • Context-Aware Evidence Head: A separate head was trained to predict these contrast ratios and suspicion labels, comparing "mask-only," "image-aware," and "context-aware" inputs.
   • Multi-Architecture Verification: To determine if FP characteristics were model-specific, the evidence head was trained on four additional frozen backbones (bare nnU-Net, bare U-Net, bare Mamba, MIGF-Mamba) across all 5 PI-CAI folds.
   • External Validation: Both the refinement head and the mechanism analysis were applied to Prostate158 without retraining to test generalizability.

Key Results

• FP Imaging Characteristics: False positives consistently exhibited raw imaging contrast closer to true cancers than to benign tissue.
  • In T2-weighted and ADC channels, FP contrast ratios aligned with the direction of true cancers (e.g., ADC: FP mean −0.56 vs. Benign 0.00; Cancer −0.31).
  • This "contrast-matching" phenomenon reproduced in 35/35 observations across five distinct backbone architectures and 105/105 observations across seven modality-perturbation scenarios.
  • The finding held in the external Prostate158 cohort, confirming it is a data-level imaging property of prostate MRI rather than a model-specific artifact.
• Specificity Refinement Performance (PI-CAI Fold-0):
  • The post-hoc refinement head increased case-level specificity from 0.469 to 0.549 (+17.2% relative) while maintaining case-level sensitivity at 0.943.
  • The PI-CAI Score improved from 0.730 to 0.741.
  • Voxel-level Dice scores remained unchanged, confirming the gain resulted from case-level FP suppression.
  • The refinement outperformed monotonic recalibration baselines (temperature/Platt scaling) by +0.038 in case-level specificity at matched sensitivity.
• Generalizability and Limitations:
  • Fold-Conditional Behavior: In 5-fold cross-validation, the specificity gain was positive in 9 of 15 observations, with a range from −22% to +28%. This indicates the refinement effect is sensitive to training-set composition.
  • External Cohort (Prostate158): Both the bare and refined models saturated near full sensitivity (≈1.0) due to domain shifts in ADC data. Consequently, specificity collapsed to near zero for both, and the McNemar test showed no significant difference (p=0.69). However, the FP contrast-matching mechanism (FPs resembling cancer) replicated successfully in this external cohort.

Significance and Claims
The paper makes two primary claims regarding the nature of AI errors and practical mitigation in prostate MRI:

1. Mechanistic Insight: Residual false positives in prostate MRI detection are not arbitrary model errors but are contrast-matched to true cancer. They share raw imaging features with malignancy, making them radiologically interpretable as cancer-resembling regions. This suggests that a non-trivial fraction of AI-flagged regions should be triaged for radiologist adjudication rather than dismissed as suppressible algorithmic noise.
2. Practical Utility: A lightweight, post-hoc refinement head can improve case-level specificity (by +17.2% on the development fold) without sacrificing sensitivity. However, the magnitude of this improvement is fold-conditional, meaning clinical deployment requires validation on the target cohort before integration.

The authors conclude that while the refinement head offers a practical, parameter-efficient method to boost specificity in-domain, the underlying FP mechanism is a fundamental property of the imaging data itself, reproducible across diverse architectures and external cohorts.