학술

Impact of MRI sequence heterogeneity on deep learning–based segmentation of brain metastases

PLOS ONE
CC BY
이 매체는 공공·자유 라이선스로 본문을 직접 표시합니다.

Figures
Abstract
Brain metastases (BMs) require accurate delineation for treatment planning, yet magnetic resonance imaging (MRI) acquisition protocols vary substantially in clinical practice. Such protocol-level heterogeneity may affect the generalizability of deep learning-based segmentation models. This study evaluated the impact of heterogeneous T1-weighted MRI protocols on BM segmentation performance using two three-dimensional architectures, 3D Attention U-Net and DUCK-Net 3D. A retrospective dataset of 925 contrast-enhanced T1-weighted MRI studies from 720 patients was analyzed. Models were trained under six sequence configurations, including multi-sequence, Black-Blood (BB), Non-Black-Blood (Non-BB), and three single-sequence cohorts. Independent test datasets consisted of three MRI cohorts: MPR Range, MPR Thick Range, and T1 SPACE 3D. Five-fold cross-validation was performed. Performance was evaluated using the Dice similarity coefficient (DSC), volumetric regression, and false discovery rate-corrected statistical comparisons. DUCK-Net 3D demonstrated more consistent performance across independent MRI cohorts, particularly under cross-sequence conditions. Multi-sequence training achieved the most stable cross-sequence performance, whereas single-sequence models showed greater sensitivity to protocol variability. Segmentation accuracy declined for small lesions (<0.4 mL) across both architectures. These findings suggest that MRI protocol heterogeneity affects segmentation performance and that sequence-diverse training may improve cross-sequence robustness in clinical settings.
Citation: Shin D-W, Lee CH, Kim HJ, Sung K, Kim YJ, Kim KG, et al. (2026) Impact of MRI sequence heterogeneity on deep learning–based segmentation of brain metastases. PLoS One 21(11): e0363055. https://doi.org/10.1371/journal.pone.0363055
Editor: Mario Tortora, Università degli Studi di Napoli Federico II: Universita degli Studi di Napoli Federico II, ITALY
Received: July 3, 2026; Accepted: November 3, 2026; Published: November 20, 2026
Copyright: © 2026 Shin et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The datasets generated and/or analyzed during the current study are not publicly available because they consist of patient imaging data and associated clinical information subject to ethical and privacy restrictions. Public sharing was not approved by the Institutional Review Board. The data were generated at Gachon University Gil Medical Center and are not deposited in a public repository. De-identified data may be made available to qualified researchers upon reasonable request and with approval from the Institutional Review Board of Gachon University Gil Medical Center. Requests for data access should be directed to the Institutional Review Board of Gachon University Gil Medical Center at irb@gilhospital.com.
Funding: This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: RS-2025-24536397). The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Brain metastases (BMs) are the most common malignant intracranial tumors, accounting for over 100,000 new cases annually in the United States [1]. Early diagnosis and timely treatment are generally associated with improved clinical outcomes and reduced healthcare burden across cancer types [2]. However, accurate detection of all metastatic lesions remains challenging in routine clinical practice because BMs must be differentiated from other intracranial pathologies and treatment-related imaging changes [3]. This challenge may be further compounded by increasing radiology workload and fatigue, which have been associated with reduced reporting performance and a higher risk of diagnostic error [4], as well as by small enhancing brain lesions that may mimic metastatic foci on contrast-enhanced imaging [5]. To address these limitations, deep learning- and artificial intelligence (AI)-based approaches have increasingly been investigated to support lesion detection and segmentation on magnetic resonance image (MRI) [6].
An MRI remains the reference imaging modality for detecting BMs [1,7]. Previous studies have explored deep learning approaches for BM detection and segmentation using a range of MRI protocols, including both single-sequence and multisequence inputs [8–10]. Reported performance has varied according to imaging protocol, model design, and study setting [9,10]. Notably, black-blood (BB) MRI has been reported to facilitate BM detection by suppressing enhancing vascular structures that may mimic metastatic lesions [11]. Kottlors et al. compared convolutional neural network (CNN) models trained on BB MRI and CE T1 MRI and found that the BB-based CNN achieved superior accuracy (92.3%) compared with CE T1-based CNN (85.5%) [12]. Park et al. trained a 3D U-Net model using BB and GRE MRI sequences, achieving 93.1% overall sensitivity, including 82.4% sensitivity for tumors smaller than 3 mm [13]. Pflüger et al. developed an AI model for automated detection and quantification, reporting a DSC of 0.78 and sensitivity of 0.81 [14]. Further studies have explored segmentation-oriented architectures with enhancements to improve robustness. Amemiya et al. enhanced the single-shot detector (SSD) model using feature-fusion techniques, boosting sensitivity to 98.4% for tumors >6 mm and 45.8% for tumors <3 mm [15]. Bousabarah et al. evaluated several U-Net variants, with the best reaching 0.82 sensitivity and 0.83 specificity [16]. Chartrand et al. introduced a CNN model with volume-aware loss and sampling, reporting 91% sensitivity and 81% precision overall, with a 58% reduction in false negatives for lesions measuring 2.5–6 mm [17]. Similarly, Hsu et al. used 3D V-Net integrating MRI and CT, achieving 0.76 mean DSC and 95% sensitivity [18]. Huang et al. applied VSS loss to DeepMedic, increasing sensitivity from 85.3% to 97.5% and reducing false positives by 44.4% [19]. Jünger et al. used DeepMedic on non-small cell lung cancer brain MRI data, reporting 85.1% sensitivity with an average of 1.5 false positives per patient [20]. Kikuchi et al. incorporated vascular suppression techniques, achieving a sensitivity of 91.7%, albeit with a high false-positive rate [21]. Rudie et al. developed a 3D U-Net model achieving 96.4% sensitivity for tumors >6 mm and 70.0% overall [22]. Xue et al. introduced BMDS Net using 3D-T1-MPRAGE scans, reaching 96% sensitivity, 99% specificity, and 0.85 DSC [23]. Yin et al. validated a multicenter model, reporting 95.8% sensitivity and 0.39 false positives per patient [24]. Zhou et al. proposed a two-stage model, MetNet, achieving 99% sensitivity for tumors >6 mm and 0.87 DSC [25].
Despite these advances, most prior studies have relied on standardized or limited MRI protocols, often using a single contrast-enhanced T1-weighted sequence for model development. In routine clinical environments, however, even T1-weighted imaging exhibits substantial variability in acquisition parameters and series descriptions, including differences in slice thickness, spatial resolution, orientation, fat suppression, and vessel signal suppression techniques. Such protocol heterogeneity may significantly affect lesion conspicuity and model performance.
Standardized imaging feature frameworks have also been developed to improve reproducibility and facilitate quantitative analysis in neuro-oncology. The VASARI framework and its recent extension, VASARI 2.0, provide structured MRI feature definitions, while machine-learning studies based on VASARI features have demonstrated the potential value of standardized imaging descriptors for computational neuro-oncology. These approaches further highlight the importance of accounting for imaging variability when developing quantitative MRI-based models [26,27].
Therefore, in the present study, we investigated how variations in MRI sequences influence model performance in the detection and segmentation of BMs. Furthermore, we systematically evaluated the impact of heterogeneous T1-weighted MRI acquisition protocols on deep learning-based brain metastasis segmentation.
Materials and methods
Study design and ethical approval
Study population and dataset composition
A total of 925 contrast-enhanced T1-weighted MRI studies from 720 patients diagnosed with brain metastases between December 2010 and January 2024 were included. The dataset was divided into a development dataset comprising 889 studies (684 patients) and an independent test dataset comprising 36 studies. All examinations from the same patient were assigned to the same dataset, and no patient was shared between the development and independent test datasets. A total of 17,209 and 547 tumors were identified in the development and independent test datasets, respectively.
Lesion burden varied widely, ranging from sub-milliliter metastases to lesions exceeding 10 mL. Small lesions (< 0.4 mL) represented a substantial proportion of tumors, enabling dedicated evaluation of small-lesion performance. Demographic characteristics and tumor distribution are summarized in Table 1.
MRI acquisition and protocol heterogeneity
All scans were contrast-enhanced T1-weighted MRI acquired under heterogeneous real-world clinical protocols using multiple 1.5T and 3T scanners (Philips and Siemens platforms). Although all images were T1-weighted, acquisition parameters varied across sequence type, slice thickness, spatial resolution, vascular signal suppression, and scanner hardware.
Three principal sequence categories were defined based on series description: MPR Range, MPR Thick Range, and T1 SPACE 3D. Sequences were further stratified into Black-Blood (BB) sequences incorporating vascular signal suppression and Non-Black-Blood (Non-BB) sequences retaining vascular enhancement. This stratification enabled evaluation of protocol-induced distribution shifts across MRI cohorts. Sequence distribution in training and test datasets is summarized in Table 2.
Tumor annotation
All enhancing lesions were manually segmented on contrast-enhanced T1-weighted MRI using commercially available region of interest (ROI) software (AVIEW). The annotation workload was divided among two radiation oncologists and two neurosurgeons. Each annotated case was subsequently reviewed by another clinician to confirm lesion identification and segmentation. The reviewed annotations were used as the reference standard for supervised training and evaluation. The annotators were not formally blinded to clinical information because the annotations were performed within the clinical workflow; however, lesion identification and segmentation were based on the contrast-enhanced T1-weighted MRI findings.
Data preprocessing
To reduce inter-protocol variability and standardize volumetric inputs, a structured preprocessing pipeline was applied.
Z-axis normalization.
Z-axis spacing was calculated from Digital Imaging and Communications in Medicine (DICOM) metadata. All volumes were standardized to a fixed 200 mm z-axis coverage (Z200). Volumes shorter than 200 mm were zero-padded, whereas volumes longer than 200 mm were symmetrically cropped. This ensured consistent volumetric depth across heterogeneous acquisitions.
Spatial resampling.
All volumes were resampled to 128 × 128 × 128 voxels to ensure compatibility with 3D convolutional neural networks.
Intensity normalization.
Intensity values were normalized to a range of 0–255 to reduce scanner-dependent intensity variation. The overall workflow of dataset preparation, training configuration, and evaluation is illustrated in Fig 1.
A total of 925 contrast-enhanced T1-weighted MRI studies were preprocessed using z-axis normalization, resampling to 128 × 128 × 128 voxels, and intensity normalization. Six training configurations (All Sequences, BB, Non-BB, MPR Range, MPR Thick Range, and T1 SPACE 3D) were constructed. 3D Attention U-Net and DUCK-Net 3D were trained using five-fold cross-validation and evaluated on three independent MRI cohorts. Performance was assessed using the Dice similarity coefficient (DSC), volumetric regression analysis, and false discovery rate (FDR)-corrected statistical comparisons.
Model architectures
Two 3D convolutional neural network architectures were evaluated.
3D Attention U-Net.
The 3D Attention U-Net employs an encoder-decoder framework with skip connections and attention gates. The architecture of the 3D Attention U-Net is shown in Fig 2. The attention mechanism enhances segmentation performance by suppressing irrelevant activations and emphasizing salient tumor regions. Batch normalization and ReLU activation functions were applied to stabilize training.
The model follows a 3D encoder-decoder architecture with attention-gated skip connections. Encoder features are extracted using repeated 3D convolutional blocks and downsampling with MaxPool3D, whereas decoder features are restored using UpSampling3D. Skip features from the encoder are refined by attention gates and concatenated with the corresponding upsampled decoder features before each decoder convolution block. Concatenation nodes indicate channel-wise feature fusion, with numbers representing the combined channel dimensions. The final segmentation map is generated using a Conv3D layer followed by sigmoid activation. Tensor sizes are denoted as channels × D³ for cubic feature maps.
DUCK-Net 3D.
DUCK-Net 3D incorporates multiscale convolutional blocks, residual connections, and dilated convolutions to improve feature extraction across spatial scales. The architecture of DUCK-Net 3D is shown in Fig 3. Residual connections facilitate gradient propagation, while the multiscale design enables robust feature learning under heterogeneous imaging conditions. These features may improve robustness under domain shift.
The model follows a 3D encoder-decoder architecture in which each resolution level is composed of a Duck block. Encoder features are progressively downsampled using stride-2 3D convolutions, and decoder features are restored using UpSampling3D. Skip connections transfer encoder features to the corresponding decoder levels, where they are concatenated with the upsampled decoder features for channel-wise feature fusion. Concatenation nodes indicate the combined channel dimensions after feature fusion. The final segmentation map is generated using a 1 × 1 × 1 convolution followed by sigmoid activation. Tensor sizes are denoted as channels × D³ for cubic feature maps.
Training strategy
Five-fold cross-validation was performed within the development dataset, and both models were trained and validated using identical fold splits. Repeated examinations from the same patient were assigned to the same fold to prevent patient-level information leakage. Six training configurations were defined to simulate clinically relevant deployment scenarios: All Sequences (multi-sequence pooled training), BB only, Non-BB only, MPR Range only, MPR Thick Range only, and T1 SPACE 3D only.
Final testing was performed on the patient-independent test dataset, which consisted of three sequence-specific cohorts: MPR Range, MPR Thick Range, and T1 SPACE 3D. No patient in the independent test dataset was represented in the development dataset. Each test cohort was evaluated separately to assess cross-sequence generalization.
To assess the potential effect of training-set size on the observed performance of multi-sequence training, additional balanced multi-sequence experiments were performed using fixed sample sizes of N = 89, N = 298, and N = 342. Cases were sampled across the available MRI sequence categories while maintaining approximately balanced representation among sequence types. Patient-grouped five-fold cross-validation was applied to these balanced datasets.
Training parameters.
Training was performed using the Adam optimizer with an initial learning rate of 1 × 10 ⁻ ⁴ and a batch size of 4 for a maximum of 200 epochs. Early stopping was applied with a patience of 30, and ReduceLROnPlateau was used with a factor of 0.1 and a patience of 5. Dice loss was used as the loss function.
Evaluation metrics.
The primary evaluation metric was the Dice similarity coefficient (DSC).
(1)where P denotes the predicted segmentation and G denotes the ground truth. DSC was calculated per volume and averaged across folds.
Tumor-level volumetric and lesion-level analysis.
To assess volumetric agreement, predicted tumor volumes were compared with manual annotations using linear regression analysis. The coefficient of determination (R²) and regression equation (y = ax + b) were reported. Subgroup analysis was performed for small tumors (< 0.4 mL).
Lesion-level performance was additionally assessed using 3D connected-component analysis and one-to-one matching between reference and predicted lesions based on spatial overlap. Lesion-level sensitivity, precision, and false positives per patient were calculated. Sensitivity was also evaluated separately for lesions <0.4 mL and ≥0.4 mL.
Statistical analysis
Fold-wise performance comparisons between 3D Attention U-Net and DUCK-Net 3D were conducted using paired two-sided t-tests. Because multiple training-test sequence combinations were evaluated, p-values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) correction. An FDR-adjusted q-value < 0.05 was considered statistically significant. Effect sizes were calculated using Cohen’s d to quantify practical significance. Mean values are reported with standard deviation. The 95% confidence intervals are provided in the supplementary tables, where applicable.
For the additional balanced multi-sequence analyses, case-level 95% confidence intervals were estimated using bootstrap resampling. Differences between balanced training conditions were further assessed using permutation testing.
Results
Segmentation performance across independent MRI cohorts
Segmentation performance across six training configurations and three independent MRI test cohorts is summarized in Table 3. Detailed fold-wise DSC statistics are provided in S1 Table.
The corresponding DSC trends are illustrated in Fig 4.
Mean DSC values are shown for each training configuration and test MRI cohort (MPR Range, MPR Thick Range, and T1 SPACE 3D). (A) 3D Attention U-Net. (B) DUCK-Net 3D. Error bars represent standard deviations across five cross-validation folds. Statistical comparisons between models were performed using paired t-tests with FDR correction.
Across most cross-sequence conditions, DUCK-Net 3D tended to show higher DSC values than 3D Attention U-Net.
Multi-sequence (All Sequences) training.
When trained on pooled multi-sequence data, DUCK-Net 3D achieved significantly higher performance on the MPR Range test cohort (82.41 ± 0.92%) than 3D Attention U-Net (80.30 ± 1.14%; FDR-adjusted q = 0.044; Cohen’s d = 1.72). Performance differences were not statistically significant for the MPR Thick Range cohort (q = 0.769). On T1 SPACE 3D, DUCK-Net 3D showed numerically higher performance, although statistical significance was not retained after FDR correction. These results indicate that multi-sequence training enhances cross-sequence robustness.
In the balanced multi-sequence analysis, mean case-level DSC increased from 0.2490 for N = 89 to 0.6711 for N = 298 and remained essentially unchanged at 0.6714 for N = 342 (Table 4). The corresponding 95% bootstrap confidence intervals were 0.2257–0.2707, 0.5904–0.7397, and 0.5901–0.7410, respectively. The difference between N = 342 and N = 298 was minimal (mean difference, 0.0003; 95% CI, −0.0123 to 0.0115; permutation p = 0.958), suggesting that performance had largely stabilized by N = 298.
BB training.
Under vascular-suppressed training, DUCK-Net 3D significantly outperformed 3D Attention U-Net on the MPR Range and T1 SPACE 3D test cohorts (both q = 0.038). Both comparisons demonstrated large effect sizes (Cohen’s d > 2). No significant difference was observed on MPR Thick Range.
Non-BB training.
For Non-BB training, DUCK-Net 3D showed significantly higher DSC on the MPR Range test cohort (q = 0.038). Differences for the other cohorts were not statistically significant after FDR correction.
Single-sequence training.
Single-sequence models exhibited greater sensitivity to domain shift. Notably, the Attention U-Net trained on T1 SPACE 3D showed marked degradation when tested on MPR Range (46.16 ± 7.05%), whereas DUCK-Net 3D maintained higher performance (59.10 ± 2.69%; q = 0.038). Similarly, among models trained on MPR Thick Range, DUCK-Net 3D significantly outperformed 3D Attention U-Net when tested on T1 SPACE 3D (q = 0.038). Across cross-domain scenarios, effect sizes were consistently large (Cohen’s d > 1), indicating clinically meaningful differences.
Qualitative segmentation comparison
Representative qualitative segmentation examples are shown in Fig 5. Models trained using multi-sequence or BB configurations showed improved lesion boundary delineation and fewer false positives than models trained on single-sequence cohorts. Failure cases were predominantly associated with small or low-contrast lesions.
(A) Original axial contrast-enhanced T1-weighted MRI. (B) Ground truth tumor segmentation mask. (C) Overlay of the original MRI and ground truth segmentation for reference. (D–I) Predicted segmentation results from DUCK-Net 3D models trained on different MRI sequence cohorts: (D) all sequences, (E) Non-BB, (F) BB, (G) MPR Range, (H) MPR Thick Range, and (I) T1 SPACE 3D. Red contours indicate predicted tumor regions. (J–L) Input DICOM images from the test set used to illustrate representative failure cases. (M–O) Corresponding segmentation results overlaid on each DICOM slice, where red indicates false-positive or misaligned segmentations and green indicates missed true lesions.
Tumor-level volumetric regression analysis
Regression analysis between predicted and manually annotated tumor volumes is presented separately for lesions ≥ 0.4 mL and < 0.4 mL. Results for lesions ≥ 0.4 mL are shown in Fig 6, and results for lesions < 0.4 mL are shown in Fig 7. Detailed regression statistics are provided in S2 Table and S3 Table, respectively.
(A) 3D Attention U-Net. (B) DUCK-Net 3D. Within each model, the six training configurations are shown in the following order: All sequences, Non-BB, BB, MPR Range, MPR Thick Range, and T1 SPACE 3D. Lesion-level predictions from all five cross-validation folds were pooled for analysis. Each subplot includes results from three independent MRI test cohorts: TEST MPR Range, TEST MPR Thick Range, and TEST T1 SPACE 3D. Colored symbols and regression lines indicate the individual test cohorts.
(A) 3D Attention U-Net. (B) DUCK-Net 3D. Within each model, the six training configurations are shown in the following order: All sequences, Non-BB, BB, MPR Range, MPR Thick Range, and T1 SPACE 3D. Lesion-level predictions from all five cross-validation folds were pooled for analysis. Each subplot includes results from three independent MRI test cohorts: TEST MPR Range, TEST MPR Thick Range, and TEST T1 SPACE 3D. The diagonal line indicates perfect agreement.
Multi-sequence training achieved the highest volumetric agreement (R² = 0.97), followed by Non-BB training (R² = 0.95). Performance declined in single-sequence models, particularly those trained exclusively on T1 SPACE 3D. Subgroup analysis of small tumors (< 0.4 mL) showed lower R² values across all training configurations, indicating the difficulty of accurate volumetric estimation in small metastases.
Lesion-level detection performance
Lesion-level detection results are summarized in Table 5. The N = 298 model achieved a mean sensitivity of 0.551 ± 0.056, precision of 0.848 ± 0.047, and 1.56 ± 0.72 false positives per patient. Sensitivity was substantially lower for lesions <0.4 mL (0.423 ± 0.068) than for lesions ≥0.4 mL (0.884 ± 0.029). The N = 342 model showed similar performance, with a sensitivity of 0.542 ± 0.039, precision of 0.870 ± 0.033, and 1.26 ± 0.44 false positives per patient. Corresponding sensitivities were 0.413 ± 0.055 for lesions <0.4 mL and 0.878 ± 0.035 for lesions ≥0.4 mL. In contrast, the N = 89 condition showed marked instability across folds, with substantially reduced sensitivity and precision and a large number of false-positive detections.
Discussion
Cross-sequence generalization under protocol heterogeneity
This study systematically evaluated the impact of protocol-level MRI heterogeneity on deep learning-based segmentation of brain metastases. Although all datasets consisted of contrast-enhanced T1-weighted MRI, substantial variability existed in acquisition protocols, including vascular suppression techniques and sequence-specific characteristics. The primary finding was that DUCK-Net 3D demonstrated more stable performance across independent MRI cohorts than 3D Attention U-Net, particularly under cross-sequence conditions in which training and test protocols differed. When models were trained and tested on similar sequence types, performance differences were limited. However, under domain-shift scenarios, particularly when models trained on T1 SPACE 3D were tested on MPR Range images, 3D Attention U-Net showed pronounced performance degradation. In contrast, DUCK-Net 3D maintained relatively higher Dice scores and demonstrated large effect sizes in several cross-domain comparisons. These findings suggest that architectural components such as multiscale feature extraction and residual connections may contribute to improved robustness under protocol-induced distribution shifts.
Effect of multi-sequence and vascular-suppressed training
Multi-sequence (All Sequences) training achieved the most consistent cross-cohort performance. Exposure to heterogeneous imaging characteristics during training likely enabled the model to learn feature representations that were more robust to sequence variation. BB training also showed improved performance in certain cross-sequence scenarios. BB sequences suppress vascular signals and may enhance tumor conspicuity by reducing vascular enhancement artifacts. However, improvements were not uniformly observed across all test cohorts, indicating that vascular suppression alone does not guarantee robustness. Collectively, these findings suggest that heterogeneous multi-sequence training may improve cross-sequence robustness. The balanced sample analysis further indicated that this benefit was not explained solely by increasing training-set size, although the independent contribution of sequence diversity cannot be completely separated from other dataset-related factors.
Performance in small lesions
Consistent with prior literature, segmentation accuracy declined for small metastases (< 0.4 mL). Regression analysis demonstrated reduced volumetric agreement in this subgroup, and lesion-level analysis showed substantially lower sensitivity for lesions <0.4 mL than for larger lesions. In the balanced multi-sequence models, sensitivity for lesions <0.4 mL was approximately 0.42, compared with approximately 0.88 for lesions ≥0.4 mL. Small lesions present intrinsic challenges, including low contrast-to-noise ratio, partial volume effects, and increased sensitivity to inter-sequence variability. Although DUCK-Net 3D showed improved performance compared with 3D Attention U-Net in several configurations, detection of small metastases remains suboptimal. Future work may benefit from size-aware loss functions, multiscale supervision, or lesion-level training strategies to improve sensitivity in this clinically relevant subgroup.
Clinical implications
MRI acquisition protocols can vary substantially because of differences in scanner hardware, institutional preferences, and workflow constraints. Many prior deep learning studies have relied on relatively homogeneous datasets, potentially limiting robustness to protocol variation. Our findings suggest that incorporating heterogeneous MRI sequences during training may improve cross-sequence robustness. In clinical environments where MRI protocols cannot be fully standardized, training with heterogeneous multi-sequence data may therefore provide a practical strategy for improving robustness across sequence variations. However, because the present study was conducted at a single institution, these findings should not be interpreted as evidence of multicenter generalizability. External multicenter validation will be necessary to determine whether the observed robustness extends across institutions and imaging environments.
Limitations
Several limitations should be acknowledged. First, this was a retrospective single-center study. Although multiple scanners and heterogeneous protocols were included, external multicenter validation was not performed. Second, voxel-level DSC may be influenced by class imbalance, particularly for small lesions, potentially affecting performance interpretation. Third, the Z200 normalization strategy standardized volumetric depth to 200 mm. While this improved spatial consistency, alternative cropping or normalization strategies may influence performance. In addition, resampling all images to 128 × 128 × 128 voxels may have affected the representation of very small lesions. In the independent test dataset, 14 of 547 reference lesions (2.6%) were no longer represented after resampling; all of these lesions were <0.4 mL, corresponding to 3.5% of small lesions. This potential loss of very small lesions may have contributed to the lower sensitivity observed in this subgroup. Future studies should include multicenter validation, prospective testing, and architecture-level adaptations specifically targeting small-lesion detection.
Conclusion
In this study, protocol-level MRI heterogeneity significantly influenced deep learning-based brain metastasis segmentation performance. DUCK-Net 3D demonstrated improved cross-sequence robustness compared with 3D Attention U-Net, particularly under domain-shift conditions. Multi-sequence training enhanced robustness across independent MRI cohorts, whereas single-sequence models were more vulnerable to protocol variability. Despite these improvements, small metastases remain a major challenge, highlighting the need for refined detection strategies in heterogeneous clinical environments.
Supporting information
S1 Table. Fold-wise DSC statistics across training configurations and independent MRI test cohorts.
Values are presented as percentages and include mean DSC, standard deviation, and 95% confidence intervals across five folds.
https://doi.org/10.1371/journal.pone.0363055.s001
(CSV)
S2 Table. Regression statistics for lesions ≥ 0.4 mL across training configurations and independent MRI test cohorts.
Values include regression coefficients and goodness-of-fit measures for pooled five-fold analyses.
https://doi.org/10.1371/journal.pone.0363055.s002
(CSV)
S3 Table. Regression statistics for lesions < 0.4 mL across training configurations and independent MRI test cohorts.
Values include regression coefficients and goodness-of-fit measures for pooled five-fold analyses. NA indicates values that were not estimable because of insufficient lesion-level data or zero variance in the regression analysis.
https://doi.org/10.1371/journal.pone.0363055.s003
(CSV)
References
- 1. Mitchell DK, Kwon HJ, Kubica PA, Huff WX, O’Regan R, Dey M. Brain metastases: An update on the multi-disciplinary approach of clinical management. Neurochirurgie. 2022;68(1):69–85. pmid:33864773
- 2. Aguiar-Ibáñez R, Mbous YPV, Sharma S, Chawla E. Assessing the clinical, humanistic, and economic impact of early cancer diagnosis: a systematic literature review. Front Oncol. 2025;15:1546447. pmid:40177242
- 3. Ortiz de Mendivil A, Martín-Medina P, García-Cañamaque L, Jiménez-Munarriz B, Ciérvide R, Diamantopoulos J. Challenges in radiological evaluation of brain metastases, beyond progression. Radiologia (Engl Ed). 2024;66(2):166–80. pmid:38614532
- 4. Vosshenrich J, Brantner P, Cyriac J, Boll DT, Merkle EM, Heye T. Quantifying Radiology Resident Fatigue: Analysis of Preliminary Reports. Radiology. 2021 Mar;298(3):632–9.
- 5. Okamoto K, Ito J, Furusawa T, Sakai K, Watanabe M, Tokiguchi S. Small cortical infarcts mimicking metastatic tumors. Clin Imaging. 1998;22(5):333–8. pmid:9755395
- 6. Dorfner FJ, Patel JB, Kalpathy-Cramer J, Gerstner ER, Bridge CP. A review of deep learning for brain tumor analysis in MRI. NPJ Precis Oncol. 2025;9(1):2.
- 7. Wang TW, Shiao YC, Hong JS, Lee WK, Hsu MS, Cheng HM. Artificial Intelligence Detection and Segmentation Models: A Systematic Review and Meta-Analysis of Brain Tumors in Magnetic Resonance Imaging. Mayo Clinic Proceedings: Digital Health. 2024;2(1):75–91.
- 8. Grøvik E, Yi D, Iv M, Tong E, Rubin D, Zaharchuk G. Deep learning enables automatic detection and segmentation of brain metastases on multisequence MRI. J Magn Reson Imaging. 2020;51(1):175–82. pmid:31050074
- 9. Ozkara BB, Chen MM, Federau C, Karabacak M, Briere TM, Li J, et al. Deep Learning for Detecting Brain Metastases on MRI: A Systematic Review and Meta-Analysis. Cancers (Basel). 2023;15(2):334. pmid:36672286
- 10. Wang T-W, Hsu M-S, Lee W-K, Pan H-C, Yang H-C, Lee C-C, et al. Brain metastasis tumor segmentation and detection using deep learning algorithms: A systematic review and meta-analysis. Radiother Oncol. 2024;190:110007. pmid:37967585
- 11. Oh J-H, Lee KM, Kim H-G, Yoon JT, Kim EJ. Deep learning-based detection algorithm for brain metastases on black blood imaging. Sci Rep. 2022;12(1):19503. pmid:36376364
- 12. Kottlors J, Geissen S, Jendreizik H, Große Hokamp N, Fervers P, Pennig L, et al. Contrast-Enhanced Black Blood MRI Sequence Is Superior to Conventional T1 Sequence in Automated Detection of Brain Metastases by Convolutional Neural Networks. Diagnostics (Basel). 2021;11(6):1016. pmid:34206103
- 13. Park YW, Jun Y, Lee Y, Han K, An C, Ahn SS, et al. Robust performance of deep learning for automatic detection and segmentation of brain metastases using three-dimensional black-blood and three-dimensional gradient echo imaging. Eur Radiol. 2021;31(9):6686–95. pmid:33738598
- 14. Pflüger I, Wald T, Isensee F, Schell M, Meredig H, Schlamp K. Neurooncol Adv. 2022;4(1):vdac138.
- 15. Amemiya S, Takao H, Kato S, Yamashita H, Sakamoto N, Abe O. Feature-fusion improves MRI single-shot deep learning detection of small brain metastases. J Neuroimaging. 2022;32(1):111–9. pmid:34388855
- 16. Bousabarah K, Ruge M, Brand JS, Hoevels M, Rueß D, Borggrefe J, et al. Deep convolutional neural networks for automated segmentation of brain metastases trained on clinical data. Radiat Oncol. 2020;15(1):87.
- 17. Chartrand G, Emiliani RD, Pawlowski SA, Markel DA, Bahig H, Cengarle-Samak A. Automated detection of brain metastases on T1-weighted MRI using a convolutional neural network: Impact of volume aware loss and sampling strategy. J Magn Reson Imaging. 2022;56(6):1885–98.
- 18. Hsu DG, Ballangrud Å, Shamseddine A, Deasy JO, Veeraraghavan H, Cervino L, et al. Automatic segmentation of brain metastases using T1 magnetic resonance and computed tomography images. Phys Med Biol. 2021;66(17):10.1088/1361-6560/ac1835. pmid:34315148
- 19. Huang Y, Bert C, Sommer P, Frey B, Gaipl U, Distel LV, et al. Deep learning for brain metastasis detection and segmentation in longitudinal MRI data. Med Phys. 2022;49(9):5773–86. pmid:35833351
- 20. Jünger ST, Hoyer UCI, Schaufler D, Laukamp KR, Goertz L, Thiele F, et al. Fully automated MR detection and segmentation of brain metastases in non-small cell lung cancer using deep learning. J Magn Reson Imaging. 2021;54(5):1608–22.
- 21. Kikuchi Y, Togao O, Kikuchi K, Momosaka D, Obara M, Van Cauteren M, et al. A deep convolutional neural network-based automatic detection of brain metastases with and without blood vessel suppression. Eur Radiol. 2022;32(5):2998–3005. pmid:34993572
- 22. Rudie JD, Weiss DA, Colby JB, Rauschecker AM, Laguna B, Braunstein S, et al. Three-dimensional U-Net convolutional neural network for detection and segmentation of intracranial metastases. Radiol Artif Intell. 2021;3(3):e200204.
- 23. Xue J, Wang B, Ming Y, Liu X, Jiang Z, Wang C, et al. Deep learning-based detection and segmentation-assisted management of brain metastases. Neuro Oncol. 2020;22(4):505–14. pmid:31867599
- 24. Yin S, Luo X, Yang Y, Shao Y, Ma L, Lin C, et al. Development and validation of a deep-learning model for detecting brain metastases on 3D post-contrast MRI: a multi-center multi-reader evaluation study. Neuro Oncol. 2022;24(9):1559–70. pmid:35100427
- 25. Zhou Z, Sanders JW, Johnson JM, Gule-Monroe M, Chen M, Briere TM, et al. MetNet: Computer-aided segmentation of brain metastases in post-contrast T1-weighted magnetic resonance imaging. Radiother Oncol. 2020;153:189–96. pmid:32937104
- 26. Negro A, Gemini L, Tortora M, Pace G, Iaccarino R, Marchese M, et al. VASARI 2.0: A new updated MRI VASARI lexicon to predict grading and IDH status in brain glioma. Front Oncol. 2024;14:1449982.
- 27. Gemini L, Tortora M, Giordano P, Prudente ME, Villa A, Vargas O. Vasari Scoring System in Discerning between Different Degrees of Glioma and IDH Status Prediction: A possible machine learning Application?. J Imaging. 2023;9(4):75.

관련 뉴스