ProstNFound+: μUS Prostate Cancer Detection Model

• ProstNFound+ is a μUS-based medical model designed for prostate cancer detection and risk stratification by integrating clinical data and imaging.
• It leverages a MedSAM backbone with adapter-based parameter-efficient tuning to produce interpretable heatmaps and scalar csPCa scores for targeted biopsy guidance.
• The system demonstrates robust generalization across retrospective and prospective cohorts, achieving competitive AUROC and sensitivity in real-world clinical evaluations.

ProstNFound+ is a medical foundation model system specifically adapted for prostate cancer (PCa) detection in high-frequency micro-ultrasound (μUS). The system extends the ProstNFound line by incorporating medical vision foundation models, adapter-based parameter-efficient tuning, and clinical-context prompting to achieve generalizable and interpretable lesion detection and risk stratification in both retrospective and prospective clinical settings (Wilson et al., 30 Oct 2025).

1. System Overview and Clinical Motivation

ProstNFound+ is engineered to address the limitations of existing μUS-based PCa diagnostic workflows, which depend on operator-driven, subjective visual scoring systems such as PRI-MUS for μUS or PI-RADS for mpMRI. Traditional deep learning models for μUS were typically trained from scratch, used small datasets, and were limited to cancer-vs-non-cancer discrimination without leveraging context or providing robust localization. ProstNFound+ aims to deliver two principal outputs simultaneously:

• a localized, interpretable heatmap highlighting regions suspicious for csPCa (clinically significant cancer, ISUP Grade Group ≥ 3)
• a scalar risk score representing the likelihood of csPCa for each biopsy core

This system is intended to provide a scalable, operator-independent alternative to current expert-based scoring protocols and to facilitate prospective deployment by demonstrating robust generalization to temporally and geographically distinct cohorts.

2. Architecture and Adaptation Mechanisms

The backbone of ProstNFound+ is derived from MedSAM, which itself is a vision transformer-based medical foundation model with approximately 90 million parameters in the image encoder. ProstNFound+ leverages the MedSAM encoder and mask decoder and adapts them for μUS-specific cancer detection using parameter-efficient adapter tuning. Adapters—small, trainable modules inserted into the foundation model—are optimized while the majority of pretrained weights remain fixed, minimizing overfitting risk in low-data settings.

A key component is a custom prompt encoder for integrating clinical biomarkers. This auxiliary network encodes patient age, prostate-specific antigen (PSA), and PSA density (PSAD) into 256-dimensional embeddings, which serve as contextual prompts for the mask decoder. During joint training, the prompt embeddings, image features, and context are fused to generate spatial heatmaps and risk scores. The mask decoder, based on attention and transposed convolutions (~6M parameters), generates a 256×256 cancer likelihood map per image.

The architecture further includes a multi-head output: one head produces the heatmap for spatial localization, and another generates a global csPCa probability for each input image. This dual-head setup aligns with the model's dual clinical objectives—decision support for targeted biopsy and global risk stratification.

3. Training Objectives, Loss Formulations, and Data Regimes

Training is conducted on a retrospective multicenter dataset (693 patients, 6607 biopsy cores) with systematic and targeted biopsy under μUS guidance. Ground truth is provided at the biopsy core level (benign, isPCa, csPCa). The loss function is a multi-task sum:

• For the risk head: cross-entropy

$$\mathcal{L}_\text{csPCa} = H(Y_\text{csPCa}, \hat{Y}_\text{csPCa})$$

• For the heatmap head: biopsy region-averaged cross-entropy, where predicted core involvement is the mean activation over the needle region:

$$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$

$\mathcal{L}_\text{hmap} = H(I, \hat{I})$

• Combined objective:

$$\mathcal{L} = \mathcal{L}_\text{csPCa} + \mathcal{L}_\text{hmap}$$

Retrospective training employs five-fold cross-validation with separate subject splits. Ablation studies confirm the benefit of prompt-based conditioning and the multi-head loss compared to either no prompt or single-head alternatives.

For prospective evaluation, the model is tested on data acquired five years later from a new clinical site (77 subjects, 1040 biopsy cores), with data preprocessing and normalization carried forward from the retrospective phase. Notably, PSAD was unavailable for the prospective cohort, so only age and PSA were used as prompts. The model is held fixed during prospective assessment, ensuring a true unseen validation.

4. Evaluation Schemes, Baseline Comparisons, and Quantitative Results

The ProstNFound+ framework is benchmarked against classical and contemporary baselines, including Patch-ResNet, MicroSegNet-FT, SAM/MedSAM-UNETR, and Cinepro. The principal external comparator in clinical practice is the human-scored PRI-MUS protocol for μUS as well as PI-RADS for mpMRI, both on ordinal (1–5) scales. To permit direct comparison, ProstNFound+ discretizes its continuous risk scores into 1–5 bins by histogram matching to the empirical PRI-MUS score distribution.

Retrospective Validation

On the retrospective validation:

• ProstNFound+ achieves a csPCa AUROC of $77.5 \pm 3.7$ (mean ± SD), outperforming other models such as MicroSegNet-FT ($73.9 \pm 0.9$), SAM-UNETR ($71.4 \pm 4.0$), and Patch-ResNet ($68.0 \pm 3.6$).
• Sensitivity at 60% specificity is $79.3 \pm 5.5$, a modest improvement over the original ProstNFound and other baselines.

Prospective Performance

On the prospective, temporally/geographically distinct dataset:

• On all biopsy samples ($n=1031$), the csPCa AUROC is $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$0 for ProstNFound+ versus $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$1 for PRI-MUS. Sensitivity/specificity for ProstNFound+ are $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$2, for PRI-MUS $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$3.
• On the subset with PI-RADS scores ($$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$4), ProstNFound+ csPCa AUROC is $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$5, close to PI-RADS ($$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$6) and within 4.5 points of PRI-MUS ($$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$7). Sensitivity/specificity for ProstNFound+ are $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$8, for PI-RADS $$\hat{I} = \frac{1}{|\mathcal{N}|} \sum_{i,j \in \mathcal{N}} X_{i,j}$$9.

No significant degradation in performance is observed between retrospective and prospective settings, which supports the model's robustness to domain and temporal drift.

Interpretability and Localization Analysis

ProstNFound+ produces heatmaps overlaid on μUS images, enabling visual correlation with needle region and clinical findings. The model's activations correlate with cancer involvement in the core, and the system demonstrates stronger performance with increasing lesion burden (exceeding PRI-MUS when tumor involvement exceeds 40%).

5. Methodological Innovations and Design Considerations

Several methodological contributions underpin ProstNFound+:

• Adapter Tuning: Reduces overfitting and resource demands by restricting training to adapter modules, optimal when data are limited and domain shift is substantial.
• Conditional Clinical Prompting: Embedding clinical features as prompts allows the model to integrate non-image risk factors, simulating a human reader who considers patient history in image interpretation.
• Multi-Head Output: Simultaneous spatial localization and global risk assessment expand model utility for both targeted diagnostic intervention and risk stratification.
• Weak Localization Supervision: Due to lack of pixel-level annotations, the heatmap decoder is weakly guided via core-level involvement, a practical design in datasets where only biopsy results are available.

The system is computationally efficient, operating at 27.6 FPS on GPU and 2.63 FPS on CPU, facilitating near real-time evaluation.

6. Clinical Implications, Limitations, and Interpretation

ProstNFound+ is positioned as a scalable, interpretable adjunct or partial alternative to human-centric visual scoring. Its main advantages are operator independence, interpretability (through spatial heatmaps), and generalization across sites and time, as evidenced by sustained performance in both retrospective and prospective evaluation.

Major limitations include:

• Weak localization supervision due to absence of pixelwise tumor labels, constraining precise heatmap localization.
• Restricted to 2D slices rather than full 3D prostate context, whereas human experts benefit from full-gland visual scanning.
• Although generalization is robust, sensitivity remains below PRI-MUS for smaller/less conspicuous lesions in the prospective set.
• Direct measurement of the system's impact on real-time biopsy guidance and downstream patient outcomes remains to be established in further clinical trials.

A plausible implication is that integrating pixel-level annotations, expanding to 3D inference, or coupling with additional clinical variables could further enhance performance and clinical adoption.

7. Positioning in the Prostate Foundation Model Ecosystem

ProstNFound+ exemplifies a targeted, clinically driven adaptation of large medical vision foundation models to the prostate domain, using μUS rather than MRI. Relative to related efforts such as ProFound (volumetric mpMRI, multi-task foundation modeling (Wang et al., 4 Mar 2026)) and ProViCNet (MRI/TRUS, patch-level contrastive learning (Lee et al., 1 Feb 2025)), ProstNFound+ prioritizes real-time, interpretable, and risk-stratified prostate cancer detection from cost-effective imaging, validated in a true prospective manner.

The table below summarizes distinguishing features of key prostate foundation model efforts.

Model/System	Modality	Architecture	Supervision	Downstream Tasks	Prospective Validation
ProstNFound+	μUS	MedSAM + adapters	Weak (core-level)	Localization, csPCa risk	Yes
ProFound	mpMRI (3D)	3D MAE (ViT, Conv)	Self-supervised	11 clinical (multi-task)	No
ProViCNet	mpMRI/TRUS	3D ViT + contrastive	Patch-level	Detection, localization	No

*Editor's term: "prostate FM ecosystem" — collaborative development of foundation models tailored specifically for comprehensive prostate imaging and clinical workflows.

• ProstNFound+: A Prospective Study using Medical Foundation Models for Prostate Cancer Detection (Wilson et al., 30 Oct 2025)
• ProFound: A moderate-sized vision foundation model for multi-task prostate imaging (Wang et al., 4 Mar 2026)
• Prostate-Specific Foundation Models for Enhanced Detection of Clinically Significant Cancer (Lee et al., 1 Feb 2025)