A conditional generative model for T2-weighted prostate MRI with downstream evaluation in clinically significant cancer classificatio

Xavier Rafael-Palou; Alejandro Vergara-Richart; Ignacio Iborra Roncales; Carmen Prieto-de-la-Lastra; Andreu Antolin-Redondo; Juan Morote; Ana Jimenez-Pastor

cs.CVNov 1, 2026

A conditional generative model for T2-weighted prostate MRI with downstream evaluation in clinically significant cancer classificatio

Authors: Xavier Rafael-Palou, Alejandro Vergara-Richart, Ignacio Iborra Roncales, Carmen Prieto-de-la-Lastra, Andreu Antolin-Redondo, Juan Morote, Ana Jimenez-Pastor

Organizations: Quantitative Imaging Biomarkers in Medicine · Quantitative Imaging Biomarkers in Medicine (Quibim S.L.), Valencia, Spain · Universitat Politècnica de València, València, Spain · Universitat Polit`ecnica de Val`encia, Val`encia, Spain ·

Abstract

Scarce and unrepresentative training data limit the performance of prostate MRI prediction models. Synthetic T2-weighted (T2W) MRI may augment training data containing clinically significant prostate cancer (csPCa) and transition zone (TZ) lesions. In this retrospective, single-center, ethics-committee-approved study, we developed a conditional generative model for T2W prostate MRI extending MAISI-v2 with csPCa class, zonal, and lesion-mask-guided conditioning, incorporating a relative advantage weighted loss fine-tuned on 1,303 patients (1,654 lesions). Fidelity was assessed by incremental ablation, and downstream utility was evaluated with T2W-based csPCa classifiers across five simulated training-data scenarios, each testing 18 synthetic-data configurations. The proposed model reduced Fréchet inception distance (FID) from 13.28 (pretrained MAISI-v2) to $2.04 \pm 0.06$ internally (PI-CAI: $3.83 \pm 0.51$ ). Synthetic augmentation significantly increased AUC in all five scenarios (+0.027 to +0.088; adjusted $p = 0.016$ -- $0.043$ ), though cross-institutional augmentation of PI-CAI yielded a non-significant gain (+0.013; $p = 0.397$ ). Overall, our model improved T2W MRI fidelity, and synthetic images improved csPCa classification under simulated data constraints, although benefit did not transfer significantly across institutions.

Figures & tables

Abbreviation	Definition
ADC	Apparent diffusion coefficient
AUC	Area under the receiver operating characteristic curve
BH	Benjamini–Hochberg
CNet	ControlNet
csPCa	Clinically significant prostate cancer
CV	Cross-validation

Table 1

	Train		Validation		Test
	Lesion level	Patient level	Lesion level	Patient level	Lesion level	Patient level
Cases	1,335	1,047	159	126	160	130
csPCa $-$ (csPCa+)	855 (480)	609 (438)	112 (47)	84 (42)	106 (54)	81 (49)
Lesion PZ (TZ)	877 (458)	683 (364)	110 (49)	90 (36)	100 (60)	83 (47)
Prostate vol. (mL)	53.0 (40.0, 73.0)	53.0 (40.0, 74.0)	60.0 (42.0, 87.0)	59.5 (40.0, 84.8)	52.0 (35.0, 78.3)	52.5 (37.0, 79.0)
Spacing XY (mm)	0.625 (0.56, 0.63)	0.625 (0.56, 0.63)	0.625 (0.56, 0.63)	0.625 (0.56, 0.63)	0.625 (0.56, 0.63)	0.625 (0.56, 0.63)

Table 1: Cohort characteristics of the training, validation, and test partitions, reported at the lesion level and patient level. Continuous variables are reported as median (interquartile range, 25th–75th percentiles). csPCa

-

(csPCa+), cases without (with) clinically significant prostate cancer; Lesion PZ (TZ), cases with the lesion in the peripheral (transition) zone; FoV, field of view.

Figure 1: Study flowchart. Left, in-house cohort: exclusions, patient-level partitions for generative model development and csPCa classification, and the five training scenarios augmented with the proposed generative model. Right, PI-CAI cohort: patient-level split, 5-fold cross-validation for configuration and threshold selection, external generative fidelity evaluation, and cross-institutional classification with in-house synthetic subsets. PCa, prostate cancer; csPCa, clinically significant PCa; PZ, peripheral zone; TZ, transition zone; CV, cross-validation; R, real-only; R+S, real plus synthetic.

Figure 2: Overview of the proposed generative model, comprising three networks: variational autoencoder (VAE), latent diffusion model (LDM) and ControlNet (CNET). The ControlNet receives the multi-class zone and lesion mask, the noisy latent ( $x_{t}$ ), and the timestep ( $t$ ), modality and csPCa embedding, and injects its features into the frozen LDM U-Net, which also receives the voxel spacing; the frozen VAE decoder maps the generated latent ( $z$ ) to the synthetic T2-weighted image. csPCa, clinically significant prostate cancer.

		Scenario training set (real data)
Scenario	Constraint applied	Total	csPCa+	csPCa $-$	TZ Lesions	PZ Lesions	csPCa+ & TZ
Full data	None. All downstream training set	968	394	574	345	623	147
Low data	30% of downstream training set	290	118	172	103	187	49
csPCa+ imbal.	30% of csPCa+ lesions from downstream train	693	119	574	247	446	49
TZ imbal.	30% of TZ lesions from downstream train	727	296	431	104	623	49
csPCa+/TZ imbal.	30% of csPCa+ TZ lesions from downstream train	866	292	574	243	623	45

Table 2: Simulated training-data scenarios and characteristics of their real training sets (patient level). csPCa, clinically significant prostate cancer; PZ, peripheral zone; TZ, transition zone.

Factor	Levels	Description
Synthetic size	10%, 20%, 30%	Proportion of synthetic images relative to the real training set of each scenario.
Class proportion	Balanced, TR-ratio, TR-inv	csPCa+/csPCa $-$ ratio of synthetic subset. Balanced : equal csPCa+/csPCa $-$ . TR-ratio : same ratio as real training set. TR-inv : inverse ratio (e.g., 30:70 becomes 70:30) to counterbalance imbalance.
Case selection	Fixed, Bounded	Selection criterion from generated pool. Fixed : first N images without filtering. Bounded : uncertainty-guided, retaining images with predicted csPCa+ probability between 0.40 and 0.60.

Table 3: Design of synthetic augmentation configurations (18 configurations;

3\times 3\times 2

). csPCa, clinically significant prostate cancer; TR, real training set.

ID	Main components	N Test	FID	KID	Prec	Rec
M0 (MAISI-v2)	LDM	130	$13.28\pm 2.01$	$0.0093\pm 4.7\times 10^{-4}$	$0.302\pm 0.06$	$0.014\pm 0.01$
M1	LDM	130	$3.587\pm 0.37$	$0.0027\pm 5.2\times 10^{-4}$	$0.514\pm 0.08$	$0.287\pm 0.09$
M2	LDM + csPCa	130	$3.173\pm 0.31$	$0.0023\pm 4.4\times 10^{-4}$	$0.558\pm 0.07$	$0.315\pm 0.09$
M3	LDM + csPCa + EMA	130	$2.789\pm 0.27$	$0.0019\pm 3.8\times 10^{-4}$	$0.531\pm 0.07$	$0.354\pm 0.09$
M4	CNet + LDM (M3)	160	$2.653\pm 0.08$	$0.0019\pm 9.4\times 10^{-4}$	$0.559\pm 0.08$	$0.385\pm 0.10$
M5	CNet + LDM (M3) + relLoss	160	$2.439\pm 0.08$	$0.0017\pm 8.1\times 10^{-4}$	$0.574\pm 0.08$	$0.390\pm 0.10$

Table 4: Quantitative test-set ablation results. LDM models were trained and tested on patient-level dataset, whereas CNet-based models used the lesion-level dataset. Values are mean

\pm

SD across the three orthogonal planes. M0 is the pretrained MAISI-v2 model without fine-tuning. CNet, ControlNet; csPCa, clinically significant prostate cancer class conditioning; EMA, exponential moving average; FID, Fréchet inception distance; KID, kernel inception distance; LDM, latent diffusion model; Prec, precision; Rec, recall; relLoss, relative advantage loss; relWLoss, relative advantage weighted loss; wLoss, region-weighted loss.

Figure 3: Generative ablation results on the in-house test set: Fréchet inception distance (FID; left, log scale; values above bars are means) and precision and recall (right) for the pretrained baseline (M0), latent diffusion model (LDM) ablations (M1–M3), ControlNet and loss ablations (M4–M6), and the proposed model (M7, Ours). Error bars represent SD across the three orthogonal planes.

Figure 4: Representative axial T2-weighted slices of four cases (rows): real image (first column) and synthetic images from M3 (best LDM without mask conditioning; second column), M5 (ControlNet with relative advantage loss; third column), M6 (ControlNet with region-weighted loss; fourth column), and M7 (proposed model, ControlNet with relative advantage weighted loss; last column).

		Synthetic criteria			Test results
Scenario	Data	Sz	Prop	Sel	AUC	F1	Prec	Sens	Spec
Full data	R	–	–	–	$0.673\pm 0.01$	$0.561\pm 0.05$	$0.487\pm 0.02$	$0.669\pm 0.12$	$0.576\pm 0.06$
Full data	R+S	10	TR-inv	Bound	$0.705\pm 0.01$	$0.582\pm 0.05$	$0.485\pm 0.01$	$0.743\pm 0.15$	$0.523\pm 0.07$
Low data	R	–	–	–	$0.616\pm 0.08$	$0.552\pm 0.01$	$0.452\pm 0.04$	$0.721\pm 0.06$	$0.457\pm 0.13$
Low data	R+S	20	TR-ratio	Bound	$0.704\pm 0.02$	$0.578\pm 0.02$	$0.493\pm 0.06$	$0.739\pm 0.16$	$0.502\pm 0.26$
csPCa+ imbal.	R	–	–	–	$0.590\pm 0.02$	$0.516\pm 0.03$	$0.414\pm 0.02$	$0.700\pm 0.12$	$0.394\pm 0.15$

Table 5: csPCa test-set results of the best configurations trained with real data only (R) and with synthetic augmentation (R+S) in each scenario. Synthetic criteria apply only to R+S: Sz, synthetic size (% of the real training set); Prop, class proportion; Sel, case selection (Table 3 ). Values are mean

\pm

SD across five seeds. AUC, area under the curve; Bound, uncertainty-bounded selection; csPCa, clinically significant prostate cancer; TR-inv, inverse ratio; TR-ratio, real ratio; TZ, transition zone.

Figure 5: Effect of synthetic T2-weighted (T2W) augmentation (R+S) versus real data only (R) in the five training scenarios: test AUC (A) and F1-score (B), shown as mean and SD across five seeds with the R+S $-$ R difference above each pair, and heatmap of the R+S $-$ R differences for all test metrics (C). AUC, area under the receiver operating characteristic curve; csPCa, clinically significant prostate cancer; imbal., imbalance; TZ, transition zone.

Medical Image Synthesis

Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.

See all 317 papers

PCaPaint: Prostate Cancer Inpainting by Mitigating Shortcut Learning
Sep 29, 2026Levente Lippenszky, Hongxu Yang, Marcell Dömötör +2Synthetic Data Augmentation Latent Diffusion Models
GE HealthCare
Do Synthetic Brain MRIs Reliably Improve Tumour Classification? A StyleGAN2-ADA Class-Plane Augmentation Study on BRISC 2025
May 21, 2026José Rafael Noriega CedeñoSynthetic Data Augmentation Medical Image Classification
Faculty of Applied Sciences and Technology
Diffusion-Based Synthesis of Complete 3D Biparametric Prostate MRI
Oct 22, 2026Claudia Giardina, Montse Pardàs, Verónica VilaplanaMedical Image Synthesis Magnetic Resonance Imaging
Universitat Politècnica de Catalunya, Barcelona, Spain

Sep 29, 2026J/K move · Enter open · S save

PCaPaint: Prostate Cancer Inpainting by Mitigating Shortcut Learning

Levente Lippenszky, Hongxu Yang, Marcell Dömötör, Krisztian Koos+1

GE HealthCare

Synthetic Data Augmentation Latent Diffusion Models Magnetic Resonance Imaging Image Inpainting Medical Image Synthesis Shortcut Learning

Open paper AskPDF arXiv

The development of AI systems for tumor-specific applications is limited by the scarcity of labeled data. Synthetic tumor inpainting offers a promising approach but faces challenges for prostate cancer MRI which contains high-resolution multi-sequence data. Although methods leveraging latent diffusion models (LDMs) enable large-volume synthesis, they are prone to shortcut learning, simply reproducing the condition image created by masking the lesion region. In this work, we introduce PCaPaint, a prostate cancer inpainting method based on LDMs that explicitly addresses this failure mode. To overcome shortcut learning that compromises synthetic tumor texture, we propose a simple yet efficient conditioning strategy in which the condition image is filled with Gaussian noise, and we provide theoretical justification. In addition, we propose a novel training objective for LDM that emphasizes the error within the lesion region. Furthermore, we introduce a multi-sequence latent design, in which T2w scans and DWI&ADC scans are compressed using two separate autoencoders to preserve their distinct frequency characteristics. Extensive experiments demonstrate that the generated synthetic data improves downstream performance in prostate lesion segmentation, patient-level classification and lesion-level detection. Furthermore, our method significantly outperforms a recent state-of-the-art LDM-based tumor inpainting method both in downstream performance and in synthetic image quality.

Abstract

Medical Image Synthesis

Explore similar work

PCaPaint: Prostate Cancer Inpainting by Mitigating Shortcut Learning

Do Synthetic Brain MRIs Reliably Improve Tumour Classification? A StyleGAN2-ADA Class-Plane Augmentation Study on BRISC 2025

Diffusion-Based Synthesis of Complete 3D Biparametric Prostate MRI