Diffusion-Based Tumor Inpainting for Renal Segmentation under Clinical Data Scarcity

Ekaterina Sedykh; Salme Ussanov; Dmytro Fedorenko; Dmytro Fishman

cs.CVSep 18, 2026

Diffusion-Based Tumor Inpainting for Renal Segmentation under Clinical Data Scarcity

Authors: Ekaterina Sedykh, Salme Ussanov, Dmytro Fedorenko, Dmytro Fishman

Organizations: Institute of Computer Science, University of Tartu, Tartu, Estonia · Institute of Tissue Medicine and Pathology, University of Bern, Bern, Switzerland · Department of Digital Medicine, University of Bern, Bern, Switzerland · School of Computer Science and Centre for Intelligent Machines, McGill University, Montréal, Canada · Mila – Québec Artificial Intelligence Institute, Montréal, Canada · Better Medicine OÜ, Tartu, Estonia · STACC OÜ, Tartu, Estonia

Abstract

Deep learning segmentation of renal tumors requires large annotated datasets, yet clinical deployments typically offer only a handful of tumor-positive cases from the target site. We propose a diffusion-based inpainting framework that synthesizes anatomically plausible renal tumors within healthy CT scans, requiring no additional annotation, and provide the first systematic comparison of 2D, 2.5D, and full 3D (MAISI) synthesis strategies for this task. Training the diffusion model on public data (KiTS23, KIRC) and evaluating nnU-Net segmentation on a internal cohort across three low-data regimes, we find that 2.5D and 3D augmentation substantially reduce false positives (from $\sim$ 18--20% to $\sim$ 3--6%) while maintaining Dice, whereas 2D provides no consistent benefit. Crucially, the proposed 2.5D method matches full 3D synthesis on every metric at substantially lower computational cost, indicating that local volumetric consistency alone is sufficient for effective augmentation in data- and resource-scarce clinical settings.

Figures & tables

Figure 1: The complete framework operates in three stages. (1) A DDPM is trained on public tumor-containing CT datasets (KiTS23, KIRC) with soft-tissue windowing and normalization. Both 2D and 2.5D variants are implemented. (2) For synthesis, candidate tumor masks from KiTS are geometrically processed and placed into tumor-free control scans from internal cohort using an anatomically-constrained placement module. Hard constraints prevent anatomically impossible placements (bone overlap, non-body regions), while soft penalties encourage realistic kidney-tumor interactions. The DDPM then performs context-aware inpainting to synthesize realistic tumor texture. (3) nnU-Net is trained under various data regimes and evaluated on held-out test sets using Dice coefficient and false positive analysis.

Figure 2: Comparison of the three synthetic tumor generation strategies. 2D (ours): a single target slice is inpainted independently inside the tumor mask, with no cross-slice context. 2.5D (ours): the target slice is generated using the two preceding slices as context (without mask); the generated slice then feeds into the next iteration as context, enabling a rolling window over the volume. 3D (MAISI): the full CT volume is synthesized jointly in a single pass through a latent diffusion model.

Panel of Figure 3 — Figure 3: Inputs to the anatomically-constrained placement pipeline. (a) Multi-organ segmentation of a host control scan obtained with TotalSegmentator, providing the bone, organ, and vascular structures used as hard and soft placement constraints. (b) A representative subset of the candidate tumor library: real tumor meshes extracted from KiTS cases, from which lesions are sampled, scaled, and inserted. Tumors are shown at a common scale, illustrating the diversity of size and shape available for synthesis.

Figure 4: Segmentation performance across real-data regimes and synthetic augmentation strategies. Mean Dice score (tumor), false positive rate (% of all scans), mean FP voxels, and mean FP objects, averaged over 10 random splits with 5-fold nnU-Net ensembles. Error bars show standard deviation. FP voxel error bars are clipped at zero.

Condition	Dice	IoU	FP rate (%)	FP objects	FP voxels
Real only	$0.844\pm 0.038$	$0.732\pm 0.058$	$16.4$	$20.4\pm 11.5$	$177{,}750$
+ 2D synth	$0.774\pm 0.050$	$0.634\pm 0.067$	$17.3$	$45.0\pm 9.1$	$108{,}829$
+ 2.5D synth	$\mathbf{0.855\pm 0.058}$	$\mathbf{0.750\pm 0.085}$	$2.95$	$3.0\pm 1.6$	$58{,}848$
+ 3D synth	$0.819\pm 0.055$	$0.697\pm 0.076$	$\mathbf{2.18}$	$\mathbf{1.8\pm 1.3}$	$\mathbf{42{,}709}$

Table 1: Segmentation performance at 2.5% real training data (4 scans), averaged over 10 random splits with 5-fold ensembles. Bold denotes best per metric.

Condition	Metric	2.5% (4 scans)	5% (8 scans)	10% (16 scans)
Real only	Dice	$0.844\pm 0.038$	$0.888\pm 0.028$	$0.904\pm 0.033$
Real only	FP rate	$16.4\%$	$19.9\%$	$18.6\%$
+ 2D synth	Dice	$0.774\pm 0.050$	$0.877\pm 0.028$	$0.901\pm 0.020$
+ 2D synth	FP rate	$17.3\%$	$10.0\%$	$15.9\%$
+ 2.5D synth	Dice	$0.855\pm 0.058$	$0.924\pm 0.019$	$0.939\pm 0.021$
+ 2.5D synth	FP rate	$2.95\%$	$5.13\%$	$6.41\%$

Table 2: Mean Dice and FP rate across all regimes and augmentation conditions (10 splits, 5-fold ensembles).

Comparison	2.5% real	5% real	10% real
Real vs. 2D	$\mathbf{0.000}$	$0.294$	$0.840$
Real vs. 2.5D	$0.485$	$\mathbf{0.005}$	$\mathbf{0.006}$
Real vs. 3D	$0.221$	$\mathbf{0.035}$	$\mathbf{0.035}$
2D vs. 2.5D	$\mathbf{0.000}$	$\mathbf{0.003}$	$\mathbf{0.000}$
2D vs. 3D	$\mathbf{0.031}$	$\mathbf{0.005}$	$\mathbf{0.010}$
2.5D vs. 3D	$0.103$	$0.365$	$0.995$

Table 3: Pairwise significance (

p

-values, paired

t

-test) for Dice between augmentation conditions, per regime.

Comparison	2.5%	5%	10%
Real vs. +2D	.0004 (Real)	.391	.878
Real vs. +2.5D	.545	.010 (2.5D)	.011 (2.5D)
Real vs. +3D	.306	.053	.053
+2D vs. +2.5D	$<$ .001 (2.5D)	.007 (2.5D)	$<$ .001 (2.5D)
+2D vs. +3D	.051	.011 (3D)	.018 (3D)
+2.5D vs. +3D	.149	.438	.995

Table 4: Pairwise BH-adjusted

p

-values (paired

t

-test) per regime. Bold =

p<0.05

; winner shown in parentheses where significant. BH-adjusted over 36 tests (6 comparisons

\times

3 regimes

\times

2 metrics).

Figure 5: Tumor crops from five consecutive axial slices generated by each synthesis strategy. Each row shows the same lesion across adjacent slices; outlines indicate the tumor boundary. The 2D method exhibits pronounced changes in texture and contrast between consecutive slices, whereas the context-aware methods (2.5D, 3D) maintain a consistent tumor appearance throughout the volume.

Method	FID vs. KiTS+KIRC	FID vs. TUH real
2D (ours)	$77.13$	$66.65$
2.5D (ours)	$\mathbf{76.20}$	$\mathbf{65.97}$
3D (MAISI)	$80.02$	$80.75$

Table 5: Tumor-region FID for each synthesis strategy against two references: the diffusion training distribution (KiTS23+KIRC) and the real internal TUH cohort tumor distribution (

n=191

). FID computed on fixed-size axial tumor crops; lower is better. Values are relative indicators (ImageNet-pretrained Inception features on grayscale CT;

n_{\text{crops}}

reported below).

Explore similar work

FOSCU: Feasibility of Synthetic MRI Generation via Duo-Diffusion Models for Enhancement of 3D U-Nets in Hepatic Segmentation
Mar 31, 2026Youngung Han, Kyeonghun Kim, Seoyoung Ju +8Medical Image Segmentation Synthetic Data Augmentation
Seoul National University · OUTTA · Sangmyung University +4
Report Supervision
Aug 27, 2026Pedro R. A. S. Bassia, Wenxuan Li, Jakob Wasserthal +10Medical Image Segmentation Weakly Supervised Image Segmentation
Department of Computer Science, Johns Hopkins University, Baltimore, MD, USA · Clinic of Radiology and Nuclear Medicine, University Hospital Basel, Basel, Switzerland · Department of Computer Science, Stanford University, Stanford, CA, USA +5
PCaPaint: Prostate Cancer Inpainting by Mitigating Shortcut Learning
Sep 29, 2026Levente Lippenszky, Hongxu Yang, Marcell Dömötör +2Synthetic Data Augmentation Latent Diffusion Models
GE HealthCare

Mar 31, 2026J/K move · Enter open · S save

FOSCU: Feasibility of Synthetic MRI Generation via Duo-Diffusion Models for Enhancement of 3D U-Nets in Hepatic Segmentation

Youngung Han, Kyeonghun Kim, Seoyoung Ju, Yeonju Jean+7

Seoul National University · OUTTA · Sangmyung University+4

Medical Image Segmentation Synthetic Data Augmentation 3D Medical Image Segmentation Medical Image Synthesis Conditional Diffusion Models

Open paper AskPDF arXiv

Medical image segmentation faces fundamental challenges including restricted access, costly annotation, and data shortage to clinical datasets through Picture Archiving and Communication Systems (PACS). These systemic barriers significantly impede the development of robust segmentation algorithms. To address these challenges, we propose FOSCU, which integrates Duo-Diffusion, a 3D latent diffusion model with ControlNet that simultaneously generates high-resolution, anatomically realistic synthetic MRI volumes and corresponding segmentation labels, and an enhanced 3D U-Net training pipeline. Duo-Diffusion employs segmentation-conditioned diffusion to ensure spatial consistency and precise anatomical detail in the generated data. Experimental evaluation on 720 abdominal MRI scans shows that models trained with combined real and synthetic data yield a mean Dice score gain of 0.67% over those using only real data, and achieve a 36.4% reduction in Fréchet Inception Distance (FID), reflecting enhanced image fidelity.

Abstract

Explore similar work

FOSCU: Feasibility of Synthetic MRI Generation via Duo-Diffusion Models for Enhancement of 3D U-Nets in Hepatic Segmentation

Report Supervision

PCaPaint: Prostate Cancer Inpainting by Mitigating Shortcut Learning