SmaRT: Style-Modulated Robust Test-Time Adaptation for Source-Free Cross-Domain Brain Tumor Segmentation

We propose a test-time adaptation method called SmaRT (Style-Modulated Robust Test-Time Adaptation).
This method is designed for source-free cross-domain brain tumor segmentation in MRI.
The main contributions of this work can be summarized as follows:

• Style-aware augmentation: dynamically generates diverse domain-specific appearances to alleviate scanner/protocol variations.
• Dual-EMA pseudo-label refinement: employs two teacher models with adaptive momentum, improving stability and reliability.
• Structural priors: incorporates consistency, integrity, and connectivity constraints to preserve tumor anatomy.
• Comprehensive validation: extensive experiments on BraTS-SSA and BraTS-PED datasets demonstrate superior segmentation accuracy and robustness compared to state-of-the-art SFUDA methods.

💡 Proposed method

Overview of the proposed SmaRT framework. A dual-branch design with style modulation and multi-head structural constraints enables stable test-time adaptation and preserves anatomical fidelity.

Dynamic composite data augmentation. Eleven candidate combinations from eight operations are adaptively weighted to generate diverse augmented views for robust target-domain adaptation.

⚡ Visual Comparison

Comparative visualization on the BraTS-SSA dataset. SmaRT achieves clearer tumor boundaries and fewer false predictions than competing adaptation methods, demonstrating superior cross-domain robustness.

Ablation visualization on the BraTS-SSA dataset. Without structural heads or style/augmentation modules, predictions become fragmented, whereas the full model yields coherent and stable segmentations.

Confidence calibration. Predicted probability distributions show that SmaRT yields better-calibrated outputs, improving reliability of pseudo-labels.

Uncertainty-guided refinement. SmaRT produces stable tumor boundaries across thresholds, reducing false positives and negatives compared with the NoTTA baseline.

🔧 Environment Setup

Please prepare an environment with Python 3.8, and then install the dependencies:

conda create -n SmaRT python=3.8
conda activate SmaRT
pip install -r requirements.txt

🏋️ Pre-train on Source Domain (BraTS 2024)

Run train_source.py to obtain pre-trained weights:

python train_source.py

🧪 Test-Time Adaptation in Target Domain (BraTS-SSA/PED)

⚠️ Note: Before running test.py, please check the parameter settings in the config file and update the weight path to the pre-trained checkpoint.

Run test.py to adapt and evaluate on the target domain:

python test.py

🚀 Results

Comparison of SmaRT with other state-of-the-art models on the BraTS-SSA Dataset:

Metric	Region	NoTTA	STDR	DANN	RSA	TENT	SaTTCA	DGTTA	SFDA-FSM	AIF-SFDA	ProSFDA	Ours
Dice	ET	0.5232	0.5483	0.5394	0.5527	0.6145	0.5929	0.5995	0.6118	0.6375	0.5883	0.6358
	TC	0.5153	0.5784	0.5534	0.5616	0.6010	0.5791	0.5839	0.5936	0.6092	0.5796	0.6251
	WT	0.7468	0.7505	0.7506	0.7593	0.7415	0.7195	0.7384	0.6761	0.7358	0.7412	0.7909
HD95	ET	26.85	15.19	12.66	17.66	5.50	7.36	6.98	18.40	10.30	6.79	4.80
	TC	27.04	14.88	12.50	17.69	5.78	7.61	7.38	16.82	9.27	7.05	4.93
	WT	15.22	12.29	14.20	15.80	2.79	4.49	3.95	12.27	7.35	5.36	2.62
IoU	ET	0.3847	0.4202	0.4078	0.4207	0.4732	0.4564	0.4621	0.4789	0.5044	0.4600	0.4907
	TC	0.3913	0.4648	0.4344	0.4423	0.4323	0.4305	0.4320	0.4746	0.4897	0.4477	0.4974
	WT	0.6409	0.6533	0.6399	0.6514	0.6583	0.6157	0.6397	0.5482	0.6166	0.6503	0.6731
Sensitivity	ET	0.6328	0.6202	0.6587	0.6585	0.6381	0.6397	0.6375	0.6671	0.6527	0.6318	0.6757
	TC	0.6302	0.6355	0.6577	0.6559	0.6148	0.6212	0.6183	0.6553	0.6409	0.6197	0.6694
	WT	0.6771	0.6604	0.6827	0.6826	0.6883	0.6828	0.6831	0.6791	0.6802	0.6846	0.6885

Comparison of SmaRT with other state-of-the-art models on the BraTS-PED dataset:

Metric	Region	NoTTA	STDR	DANN	RSA	TENT	SaTTCA	DGTTA	SFDA-FSM	AIF-SFDA	ProSFDA	Ours
Dice	ET	0.4232	0.4304	0.4582	0.4479	0.4086	0.4115	0.4232	0.4302	0.4421	0.4217	0.5358
	TC	0.2316	0.2642	0.2615	0.2512	0.2610	0.2755	0.2535	0.2851	0.2673	0.2417	0.3639
	WT	0.5549	0.5713	0.5687	0.5614	0.5775	0.5482	0.5837	0.4748	0.5734	0.5835	0.7265
HD95	ET	8.70	15.42	15.46	16.34	10.12	12.19	9.85	20.27	15.47	16.95	8.67
	TC	14.33	12.23	12.26	16.59	12.60	11.74	12.65	32.68	12.26	23.48	10.79
	WT	8.41	13.43	13.47	14.75	13.51	15.75	12.61	21.93	13.46	12.43	8.32
IoU	ET	0.4381	0.3462	0.3794	0.3791	0.4323	0.4059	0.4400	0.3418	0.3826	0.4084	0.5183
	TC	0.1742	0.2007	0.2062	0.2015	0.2009	0.2106	0.1934	0.2071	0.2097	0.1868	0.2672
	WT	0.4944	0.4493	0.4531	0.4287	0.4594	0.4107	0.4744	0.3514	0.4576	0.4986	0.6019
Sensitivity	ET	0.6242	0.6554	0.6497	0.6583	0.6464	0.6502	0.6437	0.6747	0.6453	0.6311	0.6925
	TC	0.5410	0.5493	0.5736	0.5715	0.5486	0.5521	0.5464	0.5534	0.5772	0.5448	0.6137
	WT	0.6341	0.6359	0.6621	0.6548	0.6686	0.6761	0.6620	0.6465	0.6589	0.6510	0.7082

Ablation study results of our proposed SmaRT on the BraTS-SSA dataset:

Metric	Region	w/o Consistency	w/o Integrity	w/o Connectivity	w/o Data Aug.	w/o Style Mod.	SmaRT
Dice	ET	0.5805	0.5597	0.5614	0.5814	0.6013	0.6358
	TC	0.5414	0.5209	0.5228	0.5429	0.5879	0.6251
	WT	0.7231	0.6994	0.7038	0.7255	0.6283	0.7909
HD95	ET	37.07	38.38	38.35	37.08	21.08	4.80
	TC	37.44	44.96	38.71	37.43	20.62	4.93
	WT	6.85	13.85	7.58	6.81	14.02	2.61
IoU	ET	0.4615	0.4423	0.4440	0.4624	0.4871	0.4907
	TC	0.4352	0.4161	0.4178	0.4367	0.4891	0.4974
	WT	0.6086	0.5829	0.5876	0.6111	0.5290	0.6731
Sensitivity	ET	0.6240	0.6188	0.6192	0.6246	0.6516	0.6757
	TC	0.6133	0.6081	0.6086	0.6140	0.6443	0.6694
	WT	0.6608	0.6512	0.6527	0.6616	0.6405	0.6885

Ablation study results of our proposed SmaRT on the BraTS-PED dataset:

Metric	Region	w/o Consistency	w/o Integrity	w/o Connectivity	w/o Data Aug.	w/o Style Mod.	SmaRT
Dice	ET	0.4697	0.4584	0.4685	0.4723	0.4119	0.5358
	TC	0.2863	0.2862	0.2874	0.2958	0.3190	0.3639
	WT	0.5941	0.6135	0.6372	0.6008	0.5644	0.7265
HD95	ET	41.37	41.14	40.62	37.52	42.03	8.66
	TC	68.84	69.00	68.45	65.04	53.13	10.78
	WT	11.68	10.75	12.80	11.03	14.03	8.31
IoU	ET	0.3886	0.3778	0.3847	0.3902	0.3323	0.5183
	TC	0.2098	0.2094	0.2100	0.2162	0.2382	0.2672
	WT	0.4856	0.5042	0.5290	0.4894	0.4569	0.6019
Sensitivity	ET	0.6685	0.6669	0.6677	0.6701	0.6795	0.6925
	TC	0.5514	0.5513	0.5513	0.5535	0.5602	0.6137
	WT	0.6694	0.6716	0.6686	0.6736	0.6732	0.7082

📁 Dataset Description

BraTS 2024 Dataset: This dataset (https://www.synapse.org/Synapse:syn59059780) contains 1,350 glioma cases, each with four MRI modalities (T1c, T1n, T2w, and T2f) and expert annotations (ET, TC, WT). In our experiments, 60 cases were randomly selected as the source-domain test set, while the remaining were used for training. Gaussian smoothing and contrast perturbation were applied during training to improve cross-domain robustness.

BraTS-SSA Dataset: This dataset (https://www.synapse.org/Synapse:syn59059780) includes 60 glioma cases from Sub-Saharan Africa, with imaging modalities and annotation categories identical to BraTS 2024. Due to low-field MRI scanners (<1.5T), motion artifacts, and low signal-to-noise ratios, the image quality is suboptimal, posing challenges for automated segmentation.

BraTS-PED Dataset: This dataset (https://www.synapse.org/Synapse:syn59059780) consists of 464 pediatric glioma cases curated by the International Pediatric Neuro-Oncology Consortium. Although the modalities are consistent with BraTS 2024, substantial heterogeneity arises from differences in anatomical location, tumor subtypes, and imaging characteristics, making segmentation more challenging.

Preprocessing steps: All images were symmetrically padded to a uniform cubic size and resampled to 128 × 128 × 128 voxels. Intensity normalization was performed within non-zero regions to reduce inter-subject variability. Finally, the four MRI modalities (T1c, T1n, T2w, FLAIR) were concatenated into a multi-channel input, ensuring consistency during both training and inference.

📝 Citation