Skip to content

ProstNFound+: A Prospective Study using Medical Foundation Models for Prostate Cancer Detection

🕒 Published (v1): 2025-10-30 17:07 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ProstNFound+ adapts MedSAM, a medical segmentation foundation model, for prostate cancer detection from micro-ultrasound (µUS) via adapter tuning, clinical metadata prompting, and a dual-head architecture producing both cancer heatmaps and clinically significant PCa (csPCa) risk scores. Its first prospective validation on data collected five years later at a new clinical site shows no performance degradation, with AUROC within ~5% of expert PRI-MUS scoring. This demonstrates that FM-based adaptation can generalize robustly to real-world clinical deployment conditions.

Problem

Deep learning models for µUS-based prostate cancer detection have been evaluated only retrospectively, leaving clinical generalizability unverified. Existing FMs lack µUS-specific feature understanding, are designed for segmentation rather than subtle tissue differentiation, and do not incorporate PCa-relevant clinical context (PSA, PSAD, age). Standard mpMRI-based PI-RADS is costly and inaccessible; µUS-based PRI-MUS requires extensive operator training and suffers from inter-observer variability.

Method

ProstNFound+ builds on MedSAM (ViT image encoder ~90M params + mask decoder ~6M params) with three adaptations:

  1. Adapter tuning: Lightweight adapters inserted into MedSAM to shift the FM toward µUS domain without full fine-tuning.
  2. Conditional prompt encoder: Three two-layer MLPs project scalar clinical biomarkers (age, PSA, PSAD) to 256-dimensional prompt embeddings, conditioning the mask decoder on patient-specific PCa risk context.
  3. Multi-head architecture: A secondary class decoder head (same attention structure as mask decoder but with a linear output layer) produces a scalar csPCa risk score alongside the spatial heatmap. Both heads are trained jointly via a two-term loss: cross-entropy on csPCa labels + involvement-based heatmap loss (mean activation in needle region vs. pathology-reported cancer involvement fraction).

The continuous csPCa score is discretized to a 1–5 risk scale via histogram matching to the PRI-MUS score distribution. Trained on 693 subjects / 6607 biopsy cores from a 5-center retrospective trial (2013–2016); prospectively tested on 77 subjects / 1040 cores from OPTIMUM trial (2021–2024, Edmonton center, blinded until training complete).

Key Contributions

  • First prospective validation of a deep learning model for µUS-based PCa detection, evaluated on data from a new site collected five years after training data.
  • Multi-head MedSAM adaptation producing simultaneous cancer localization heatmap and patient-level csPCa risk score, supporting both biopsy targeting and clinical decision-making.
  • First benchmarking of a DL model directly against PRI-MUS risk scores on a shared dataset.
  • Ablation evidence that combined age+PSA+PSAD prompting adds ~6% AUROC over no prompting, and that dual decoder heads outperform either single head alone.

Results

Retrospective (5-fold CV, n=693): - ProstNFound+ AUROC: 77.5% (best across all baselines) - vs. MedSAM-FT: 71.2%, MicroSegNet-FT: 73.9%, ProstNFound (prior): 76.6% - Sensitivity at 20%/40%/60% specificity: 95.3% / 89.3% / 79.3%

Prospective (n=77 subjects, 1031 cores): - csPCa detection AUROC: 79.1% vs. PRI-MUS 84.4% (~5% gap) - PCa (GG2+) detection AUROC: 74.2% vs. PRI-MUS 77.5% - On PI-RADS-scored subset (n=308): ProstNFound+ 87.1% vs. PI-RADS 88.1% vs. PRI-MUS 91.6% for csPCa - For cores with ≥40% cancer involvement, ProstNFound+ matches or exceeds PRI-MUS performance - Patient-level risk score 5 is highly specific for cancer (~90% of subjects with score 5 had cancer) - No performance degradation between retrospective CV and prospective test

Limitations

  • Trained on core-level biopsy labels only; pixel-level annotation unavailable, limiting localization of small lesions.
  • Only 2D sagittal frames analyzed in isolation; PRI-MUS relies on whole-gland 3D scanning, which the model cannot replicate.
  • PSAD unavailable in prospective data, so final deployed model uses only age and PSA as prompts (not the full prompt set validated in ablations).
  • Prospective cohort is small (77 subjects, single center); generalization to broader populations and hardware variants unconfirmed.
  • ~5% AUROC gap vs. expert PRI-MUS for csPCa detection; clinical biopsy guidance accuracy not directly measured.
  • No randomized trial data on downstream patient outcomes.

Relevance to Foundation Models in Medicine

ProstNFound+ is a concrete case study in domain-specific FM adaptation: it shows that a general medical segmentation FM (MedSAM) can be efficiently re-targeted to a specialized task (µUS-based cancer scoring) by combining adapters, conditional clinical prompting, and task-appropriate output heads—without large labeled datasets. The prospective validation design directly addresses a critical gap in medical FM deployment literature: demonstrating temporal and site-level generalization under realistic distribution shift. For researchers tracking medical FMs, this work establishes a methodology for embedding structured clinical metadata as FM prompts and provides a rigorous evaluation template (prospective + expert-protocol benchmark) that should become standard for clinical FM validation.