ADP-FL-MedSeg: Adaptive Differential Privacy for Federated Medical Segmentation Across Diverse Modalities

Puja Saha; Eranga Ukwatta

eess.IVApr 7, 2026

ADP-FL-MedSeg: Adaptive Differential Privacy for Federated Medical Segmentation Across Diverse Modalities

Authors: Puja Saha, Eranga Ukwatta

Organizations: College of Engineering, University of Guelph, 50 Stone Road E, Guelph, ON, Canada

Abstract

Large volumes of medical data remain underutilized because centralizing distributed data is often infeasible due to strict privacy regulations and institutional constraints. In addition, models trained in centralized settings frequently fail to generalize across clinical sites because of heterogeneity in imaging protocols and continuously evolving data distributions arising from differences in scanners, acquisition parameters, and patient populations. Federated learning offers a promising solution by enabling collaborative model training without sharing raw data. However, incorporating differential privacy into federated learning, while essential for privacy guarantees, often leads to degraded accuracy, unstable convergence, and reduced generalization. In this work, we propose an adaptive differentially private federated learning (ADP-FL) framework for medical image segmentation that dynamically adjusts privacy mechanisms to better balance the privacy-utility trade-off. The proposed approach stabilizes training, significantly improves Dice scores and segmentation boundary quality, and maintains rigorous privacy guarantees. We evaluated ADP-FL across diverse imaging modalities and segmentation tasks, including skin lesion segmentation in dermoscopic images, kidney tumor segmentation in 3D CT scans, and brain tumor segmentation in multi-parametric MRI. Compared with conventional federated learning and standard differentially private federated learning, ADP-FL consistently achieves higher accuracy, improved boundary delineation, faster convergence, and greater training stability, with performance approaching that of non-private federated learning under the same privacy budgets. These results demonstrate the practical viability of ADP-FL for high-performance, privacy-preserving medical image segmentation in real-world federated settings.

Figures & tables

Figure 1: Illustrative simulation showing how static differential privacy mechanism fail to adapt with evolving gradient distribution

Figure 2: Workflow overview of the proposed ADP-FL-MedSeg framework.

Method	NP-FL (N=500)	DP-FL (N=500)	ADP-FL (N=500)
DSC (%)	93.12 ± 0.04 / 95.18 ± 6.51	85.38 ± 0.04 / 86.88 ± 14.38	92.81 ± 0.06 / 96.27 ± 6.93

Table 1: Comparative performance (DSC %) for NP-FL, DP-FL, and ADP-FL. Results are reported as mean

\pm

std across three runs / mean

\pm

std across patients for the best model update.

Panel of Figure 3 — Figure 3: HAM10K results: (Left) DSC convergence trajectory for site-5. (Right) Qualitative segmentation comparisons between methods.

Method	KTC (N=50)	Tumor+Cyst (N=50)	Tumor (N=50)
NP-FL	94.92 ± 0.36 / 95.22 ± 1.92	81.00 ± 1.03 / 81.93 ± 17.60	79.79 ± 1.06 / 80.94 ± 18.18
DP-FL	91.94 ± 0.27 / 91.72 ± 5.89	69.08 ± 0.98 / 68.00 ± 23.85	66.29 ± 1.09 / 65.34 ± 26.38
ADP-FL	93.63 ± 0.60 / 93.96 ± 2.80	77.26 ± 0.60 / 78.96 ± 19.95	72.84 ± 2.25 / 76.09 ± 22.71

Table 2: KiTS23 Dice Similarity Coefficient (DSC %). Results are reported as Mean across three runs / Patient-level Mean for the best run .

Panel of Figure 4 — Figure 4: KiTS23: (Left) Convergence graph for site-3. (Right) Qualitative segmentation masks for Patient case 00581.

Method	ET (N=165)	NETC (N=92)	SNFH (N=210)	RC (N=179)	TC (N=165)	WT (N=210)
NP-FL	77.56 ± 0.21	57.64 ± 0.56	87.28 ± 0.06	76.86 ± 0.74	76.46 ± 0.19	88.62 ± 0.06
DP-FL	56.77 ± 0.33	24.69 ± 0.78	72.89 ± 0.40	55.01 ± 0.07	56.22 ± 0.77	76.72 ± 0.48
ADP-FL	72.27 ± 0.22	42.05 ± 4.60	85.18 ± 0.13	72.04 ± 0.22	70.97 ± 0.24	86.78 ± 0.06

Table 3: BraTS24 DSC (mean ± std) across three runs for NP-FL, DP-FL, and ADP-FL.

Method	ET (N=165)	NETC (N=92)	SNFH (N=210)	RC (N=179)	TC (N=165)	WT (N=210)
NP-FL	77.85 ± 23.04	58.38 ± 29.41	87.36 ± 13.90	77.47 ± 24.97	76.72 ± 23.40	88.71 ± 12.22
DP-FL	57.51 ± 26.64	25.97 ± 24.69	73.70 ± 18.68	56.10 ± 29.86	57.88 ± 26.80	76.48 ± 17.02
ADP-FL	72.06 ± 24.75	46.27 ± 29.50	85.35 ± 15.78	72.31 ± 27.03	71.29 ± 24.57	86.86 ± 13.96

Table 4: BraTS24 DSC (mean ± std) across patients for the best run for NP-FL, DP-FL, and ADP-FL.

Panel of Figure 5 — Figure 5: BraTS24: (Left) Convergence graph of site-4. (Right) Qualitative segmentation masks.

Metric (%)	70 th	75 th	80 th	85 th	90 th	95 th
Mean Across Runs	$92.37\pm 0.14$	$92.57\pm 0.20$	$92.71\pm 0.14$	$92.61\pm 0.02$	$92.61\pm 0.08$	$92.81\pm 0.06$
Mean Across Patients	$95.67\pm 8.11$	$96.06\pm 7.62$	$95.99\pm 7.87$	$96.24\pm 7.37$	$96.26\pm 7.29$	$96.27\pm 6.93$

Table 5: Impact of clipping percentiles on HAM10K DSC performance. Results highlight inter-run stability and intra-patient reliability.

Panel of Figure 6 — Figure 6: HAM10K sensitivity analysis: (Left) DSC convergence trajectories for various percentiles at site-5. (Right) Qualitative comparison demonstrating visual invariance to percentile shifts.

* KTC: Kidney + Tumor + Cyst
Class	70 th	75 th	80 th	85 th	90 th	95 th
KTC *	94.51 ± 0.40	93.64 ± 1.48	93.27 ± 1.32	93.97 ± 0.38	93.69 ± 0.27	93.63 ± 0.60
Tumor+Cyst	60.64 ± 18.51	66.82 ± 8.72	70.75 ± 6.21	77.39 ± 0.49	77.36 ± 1.40	77.26 ± 0.60
Tumor	66.89 ± 4.25	64.55 ± 5.05	70.39 ± 5.64	75.28 ± 1.02	74.62 ± 1.31	72.84 ± 2.25

Table 6: KiTS23 DSC (mean ± std) across three runs for different clipping percentiles.

* KTC: Kidney + Tumor + Cyst
Class	70 th	75 th	80 th	85 th	90 th	95 th
KTC *	92.15 ± 6.19	91.63 ± 6.18	95.02 ± 2.07	94.65 ± 2.78	94.78 ± 2.52	93.96 ± 2.80
Tumor + Cyst	70.21 ± 20.82	68.35 ± 22.82	74.23 ± 26.07	64.36 ± 30.89	77.27 ± 22.51	78.96 ± 19.95
Tumor	67.49 ± 23.36	65.33 ± 24.98	72.13 ± 27.53	63.07 ± 31.01	75.25 ± 24.61	76.09 ± 22.71

Table 7: KiTS23 patient-level DSC (mean ± std) for the best run across clipping percentiles.

Panel of Figure 7 — Figure 7: KiTS23 sensitivity analysis: (Left) DSC convergence trajectories for various percentiles at site-3. (Right) Qualitative comparison demonstrating visual invariance to percentile shifts.

Class	70 th	75 th	80 th	85 th	90 th	95 th
ET	47.43 ± 5.63	44.59 ± 5.47	55.82 ± 13.40	66.79 ± 2.92	69.19 ± 1.44	72.27 ± 0.22
NETC	18.23 ± 1.28	10.69 ± 5.72	26.32 ± 11.43	26.54 ± 11.62	38.24 ± 9.04	42.05 ± 4.60
SNFH	75.77 ± 9.95	81.44 ± 2.91	85.02 ± 0.63	84.96 ± 0.84	85.03 ± 0.27	85.18 ± 0.13
RC	53.20 ± 7.07	59.29 ± 8.98	59.54 ± 9.09	60.87 ± 12.12	67.63 ± 3.17	72.04 ± 0.22
TC	45.51 ± 13.98	36.69 ± 21.02	55.54 ± 7.53	67.98 ± 0.74	67.72 ± 1.25	70.97 ± 0.24
WT	83.16 ± 3.28	86.12 ± 0.42	86.41 ± 0.32	86.39 ± 0.28	86.63 ± 0.18	86.78 ± 0.06

Table 8: BraTS24 DSC (mean ± std) across three runs for different clipping percentiles.

Class	70 th	75 th	80 th	85 th	90 th	95 th
ET	57.03 ± 26.15	55.59 ± 25.27	68.04 ± 25.96	63.18 ± 26.55	47.52 ± 27.50	72.06 ± 24.75
NETC	25.29 ± 24.74	25.35 ± 24.86	16.96 ± 20.30	32.66 ± 27.67	33.84 ± 26.64	46.27 ± 29.50
SNFH	73.85 ± 18.25	72.62 ± 18.11	85.71 ± 15.54	85.55 ± 15.05	84.14 ± 17.24	85.35 ± 15.78
RC	56.33 ± 29.49	55.72 ± 29.44	70.47 ± 28.26	66.24 ± 29.60	68.07 ± 28.87	72.31 ± 27.03
TC	57.88 ± 27.22	56.86 ± 27.05	68.64 ± 26.55	68.69 ± 25.67	66.40 ± 27.09	71.29 ± 24.57
WT	77.29 ± 15.68	76.11 ± 15.73	86.57 ± 14.34	86.66 ± 13.77	86.64 ± 14.66	86.86 ± 13.96

Table 9: BraTS24 DSC (mean ± std) across patients for different clipping percentiles.

Panel of Figure 8 — Figure 8: BraTS24 sensitivity analysis: (Left) convergence graph of site-4. (Right) Qualitative results.

Explore similar work

Aug 1, 2026

Similarity Weighted Aggregation with Global Differential Privacy for Federated Brain Lesion Segmentation

Federated Learning (FL) enables collaborative training of machine learning models across multiple institutions without sharing sensitive data, making it particularly suitable for medical imaging applications. However, heterogeneous data distributions across institutions and potential information leakage through model updates remain important challenges. In this work, we propose DP-SimAgg, a privacy-preserving federated learning framework that integrates similarity-weighted aggregation with a server-side differential privacy mechanism. The proposed method applies L2 clipping to bound collaborator updates, computes similarity-based aggregation weights to mitigate the effects of non-IID data distributions, and injects calibrated Gaussian noise at the central server, providing per-round privacy guarantees under the assumed sensitivity bound. The framework is implemented using Intel's OpenFL platform and evaluated on the FeTS 2022 dataset consisting of 1251 multi-modal MRI scans for brain tumor segmentation. Experimental results demonstrate that DP-SimAgg maintains competitive segmentation performance while providing privacy protection. Under a strict per-round privacy budget (epsilon = 1, cumulative epsilon_total = 20 over 20 rounds), the method achieves Dice scores of 0.6357, 0.5305, and 0.5274 for the enhancing tumor (ET), tumor core (TC), and whole tumor (WT) regions, respectively. With a more relaxed per-round budget (epsilon = 10, cumulative epsilon_total = 200), performance approaches that of the non-private baseline while incorporating a central Gaussian mechanism with per-round (epsilon, delta)-DP accounting under the assumed sensitivity bound. These results highlight the potential of DP-SimAgg for enabling privacy-preserving collaborative learning in medical imaging applications.

Muhammad Irfan Khan, Eero Lehtonen, Joni Obradovic +4

Turku University of Applied Sciences, Turku 20520, Finland

Federated Learning Aggregation Non-IID Federated Learning Differential Privacy Brain Tumor Segmentation Federated Learning

Ask PDF

Mar 12, 2026

FL-MedSegBench: A Comprehensive Benchmark for Federated Learning on Medical Image Segmentation

Federated learning (FL) offers a privacy-preserving paradigm for collaborative medical image analysis without sharing raw data. However, the absence of standardized benchmarks for medical image segmentation hinders fair and comprehensive evaluation of FL methods. To address this gap, we introduce FL-MedSegBench, the first comprehensive benchmark for federated learning on medical image segmentation. Our benchmark encompasses nine segmentation tasks across ten imaging modalities, covering both 2D and 3D formats with realistic clinical heterogeneity. We systematically evaluate eight generic FL (gFL) and five personalized FL (pFL) methods across multiple dimensions: segmentation accuracy, fairness, communication efficiency, convergence behavior, and generalization to unseen domains. Extensive experiments reveal several key insights: (i) pFL methods, particularly those with client-specific batch normalization (e.g., FedBN), consistently outperform generic approaches; (ii) No single method universally dominates, with performance being dataset-dependent; (iii) Communication frequency analysis shows normalization-based personalization methods exhibit remarkable robustness to reduced communication frequency; (iv) Fairness evaluation identifies methods like Ditto and FedRDN that protect underperforming clients; (v) A method's generalization to unseen domains is strongly tied to its ability to perform well across participating clients. We will release an open-source toolkit to foster reproducible research and accelerate clinically applicable FL solutions, providing empirically grounded guidelines for real-world clinical deployment. The source code is available at https://github.com/meiluzhu/FL-MedSegBench.

Meilu Zhu, Zhiwei Wang, Axiu Mao +4

Department of Electrical and Computer Engineering, The University of Hong Kong, Hong Kong, China · School of Communication Engineering, Hangzhou Dianzi University, Hang Zhou, China · Department of Radiation Oncology, Stanford University, California, USA +1

Federated Learning Medical Image Benchmarks Medical Image Segmentation

Ask PDF

Oct 2, 2026

Fed-ADApt: Federated Anytime Depth Adaptation for Resource-Aware Medical Image Segmentation

Federated learning (FL) enables collaborative training of medical image segmentation models without sharing raw patient data, yet existing approaches assume a homogeneous compute budget across institutions, limiting participation of low-resource sites. We propose Fed-ADApt, a depth-adaptive federated framework for UNet-based segmentation that jointly addresses low-compute training and inference. Fed-ADApt integrates multi-depth supervision with hierarchical depth-wise aggregation, allowing each site to train according to its local compute budget while contributing to a global model that supports dynamic depth selection at deployment. We evaluated Fed-ADApt on multi-site 2D retinal fundus disc segmentation and 3D brain tumor segmentation. Across both tasks, federated collaboration substantially improves robustness under domain shift. Fed-ADApt matched the full-resource FedAvg performance in 3D and achieved competitive 2D performance with a 4.7% average Dice reduction, while reducing average inference cost by 19.5% in 3D and 34.5% in 2D and substantially reducing training cost by 98% at the most constrained sites. Importantly, Fed-ADApt enables low-resource institutions that cannot train full-capacity models to participate in federations while maintaining competitive global performance under a favorable accuracy to efficiency trade-off. By considering training and inference compute budgets, Fed-ADApt provides a practical and equitable solution for federated medical image segmentation across heterogeneous clinical and edge-enabled imaging environments.

Abhijeet Parida, Zhifan Jiang, Pooneh Roshanitabrizi +6

Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Hospital, Washington, DC, United States · Escuela Técnica Superior de Ingenieros de Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain · CIBER-BBN, Instituto de Salud Carlos III, Madrid, Spain +2

Medical Image Segmentation Heterogeneous Federated Learning Federated Learning Aggregation

Ask PDF