End-to-end 2.5D multisequence–multichannel fusion model for preoperative survival prediction in glioma: a retrospective study

Abstract

Gliomas are the most common primary malignant tumors of the central nervous system and show marked imaging heterogeneity, making accurate preoperative prediction of 24-month overall survival status important for individualized treatment planning and prognostic counseling. Summary receiver operating characteristic analyses were used to systematically compare combinations of convolutional backbone depth and input channel counts to identify an optimal sequence–channel configuration. Guided by these findings, we developed an end-to-end multisequence–multichannel fusion aggregation (EMFA) framework that integrates deep transfer learning image representations with radiomics features. A multi-instance learning (MIL) module was further incorporated to enable adaptive within-sequence weighting and cross-sequence feature aggregation, improving model interpretability. In the held-out test cohort, the EMFA framework achieved an area under the receiver operating characteristic curve (AUC) of 0.899 and an accuracy of 0.935 (93.5%), showing competitive performance relative to corresponding single-sequence baselines. The configuration analysis indicated that nine-channel inputs for T1-weighted and T2-weighted imaging and a three-channel input for contrast-enhanced T1-weighted imaging provided the best overall performance. These results suggest that EMFA offers an effective and scalable strategy for fusing multisequence magnetic resonance imaging (MRI) information for 24-month glioma survival-status prediction, supporting imaging-informatics–driven clinical decision support and potential future translation into neuroradiology workflows.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02558-y.

Introduction

Gliomas are the most common primary malignant tumors of the central nervous system [1] and exhibit marked molecular heterogeneity [2], with median overall survival typically below 21 months [3, 4]. Under the current World Health Organization (WHO) classification, diffuse gliomas are assigned grades 2–4; higher grade is associated with greater malignancy and shorter survival [5]. Accurate preoperative survival prediction is essential for individualized treatment planning and prognostic counseling.

Magnetic resonance imaging (MRI), with its superior soft-tissue contrast and spatial resolution, has become a cornerstone for the diagnosis and grading of gliomas [6–8]. Recent advances in artificial intelligence (AI) have reshaped medical image analysis, demonstrating strong potential for disease diagnosis, treatment-response prediction, prognostication, and clinical decision support [9–12]. Within this paradigm, radiomics and imaging omics have shifted imaging from qualitative visual assessment to quantitative, data-driven modeling by extracting high-throughput features that serve as putative non-invasive predictive biomarkers [13, 14], thereby improving diagnostic accuracy [15], therapeutic planning [16], and outcome prediction [17]. Deep learning (DL) further strengthens radiologic analysis through end-to-end representation learning that learns directly from data and adapts to heterogeneous datasets [18, 19]. Nevertheless, balancing spatial expressiveness with computational efficiency in high-dimensional volumetric MRI remains challenging: conventional 2D networks model through-plane context poorly, whereas fully 3D architectures, while capturing volumetric structure, are parameter-heavy and computationally expensive.

This study presents a 2.5D multisequence framework based on multi-instance learning (MIL) for preoperative survival prediction. Termed End-to-End Multisequence–Multichannel Fusion Aggregation (EMFA), the framework comprises: (i) sequence-level modeling—constructing multichannel DL backbones, transfer-learning models, and radiomics-based predictors for each MRI sequence; and (ii) ensemble-level aggregation, whereby MIL enables intra-sequence feature integration and ensemble learning fuses predictions across sequences, leveraging multimodal imaging heterogeneity for individualized prognostic assessment.

Materials and methods

Materials

Preoperative magnetic resonance imaging (MRI) and clinical data from 273 patients with glioma were included. Of these, 63 patients were consecutively collected at the Hospital (January 2006–December 2022), and 210 patients were drawn from the public BraTS 2021 dataset. Eligibility criteria were: (i) histopathologically confirmed primary glioma; (ii) completion of preoperative MRI within two weeks prior to surgery; and (iii) absence of a second primary tumor. Patients with incomplete clinical information or without corresponding MRI were excluded. For each patient, imaging comprised T1-weighted imaging (T1-w), T2-weighted imaging (T2-w), and contrast-enhanced T1-weighted imaging (CET1-w). Clinical variables—age, World Health Organization (WHO) tumor grade, and survival outcomes—were extracted from the institutional electronic medical record system or from BraTS 2021, as appropriate. All imaging and clinical data were integrated into a unified database and subjected to standardized preprocessing, annotation, and structured curation to ensure high data quality and reproducibility for subsequent model development and analysis.The patient inclusion and data screening workflow is summarized in Fig. 1. The survival endpoint definition and label coding strategy are described below.

Fig. 1.

Patient inclusion flowchart. The diagram presents inclusion and exclusion criteria and the distribution of included patients across datasets. Proportions at each stage indicate the number of patients retained relative to the immediately preceding step

Survival endpoint definition

The binary prediction endpoint was defined as 24-month overall survival status. Overall survival was calculated from the date of surgery or initial diagnosis to death or last follow-up, according to the available clinical records. Patients with overall survival of 24 months or less were assigned to the positive class, representing short-term survival or the high-risk group, whereas patients surviving for more than 24 months were assigned to the negative class, representing longer-term survival or the low-risk group. The 24-month cutoff was selected because glioma has a generally poor prognosis and this time point provides a clinically interpretable distinction between relatively short-term and longer-term survival for preoperative prognostic assessment. In the training cohort, the positive-to-negative class distribution was 49:142; in the test cohort, it was 18:64.

MRI preprocessing and ROI segmentation

All preoperative MRI scans were acquired on 1.5T or 3.0T systems (detailed acquisition parameters are provided in Supplementary File 1.A). Three sequences (T1-w, CET1-w, and T2-w) were used. Images were co-registered to the CET1-w space. To accommodate resolution variability across scanners and protocols, volumes were resampled to isotropic 1 × 1 × 1 mm3 voxels and corrected for intensity nonuniformity using N4 bias field correction. Regions of interest (ROIs) were manually delineated on MRI with ITK-SNAP (version 3.8.0; http://www.itksnap.org) [20]. CET1-w served as the primary anatomical reference for delineation because enhancing tumor margins were most clearly visualized on this sequence, while co-registered T1-w and T2-w images were reviewed simultaneously to identify non-enhancing tumor components and surrounding anatomic boundaries. The ROI was defined as the visible tumor region used for model construction. Obvious peritumoral edema on T2-w was not intentionally included when it could be separated from solid tumor; when the boundary between non-enhancing tumor and edema was ambiguous, the final boundary was determined by consensus using the co-registered CET1-w, T1-w, and T2-w images. Delineation was primarily performed by a board-certified neuroradiologist with > 10 years of experience and independently reviewed and corrected, when necessary, by a second neuroradiologist with > 15 years of experience. Both readers were blinded to clinical, pathologic, and molecular information.

EMFA framework

End-to-End Multisequence–Multichannel Fusion Aggregation (EMFA) predictive framework was proposed and implemented (Fig. 2A). The framework comprised two stages: sequence-level modeling and ensemble-level aggregation. At the sequence level, multichannel deep-learning models, transfer-learning models, and radiomics-based predictors were constructed for each MRI sequence. At the aggregation level, multi-instance learning (MIL) was used to integrate features and predictions within each sequence, and ensemble learning was applied to fuse outputs across sequences to generate the final survival prediction.

Fig. 2.

The workflow of the research. (A) Overview of the end-to-end Multisequence–Multichannel Fusion Aggregation (EMFA) model. For each glioma patient, three MRI sequences (T1-weighted imaging, contrast-enhanced T1-weighted imaging, and T2-weighted imaging) are acquired. For each sequence, multichannel deep-learning models, transfer-learning models, and radiomics-based predictors are constructed to extract complementary representations. Multi-instance learning (MIL) is then used to integrate features and predictions within each sequence, and ensemble learning combines outputs across sequences to generate the final survival prediction. (B) 2.5D multichannel deep-learning pipeline. The pipeline includes image preparation for the three sequences, ROI annotation, slice selection, construction of 2.5D multichannel inputs, training of ResNet-X transfer-learning backbones, and extraction/selection of deep transfer-learning features. (C) Schematic of the ResNet-X architectures: (a) ResNet-18, (b) ResNet-34, (c) ResNet-50, (d) ResNet-101, and (e) ResNet-152. High-resolution standalone panels for each sub-process are provided in Supplementary File 1.B

Development of the 2.5D multichannel deep transfer-learning model

Conventional deep-learning pipelines for imaging analysis often rely on the single slice containing the largest tumor cross-section, thereby neglecting contextual and three-dimensional information [21–23]. To address this, a 2.5D multichannel deep transfer-learning model was constructed to retain the computational efficiency of 2D convolutions while encoding 3D spatial context. Specifically, the slice with the largest tumor area was identified as the center layer, and its neighboring slices (± 1, ± 2, ±3, ± 4) were stacked to form inputs of varying depths (3, 5, 7, or 9 channels). In addition, axial, sagittal, and coronal views were incorporated to enhance spatial representation, yielding five distinct 2.5D multichannel input configurations.For each MRI sequence (T1-w, CET1-w, and T2-w), transfer-learning models were built for the above multichannel configurations using ResNet-18, ResNet-34, ResNet-50, ResNet-101,and ResNet-152 backbones pre-trained on the ImageNet 2012 dataset. Models were uniformly named as [Sequence]_[Channel]C_[ResNet]X, Sequence=T1, T1CE, T2; Channel = 3,5,7,9; X = 18,34,50,101,152 (e.g., T1_7C_ResNet34 denotes a seven-channel ResNet-34 model for the T1 sequence). Dropout regularization was employed to mitigate overfitting. Deep features were extracted from the avgpool layer (512 features in total) and reduced to 256 components via principal component analysis (PCA) to improve training stability and generalization. The training pipeline is illustrated in Fig. 2B–C, and further configuration details are provided in Supplementary File 1.C.

To evaluate how channel depth, network architecture, and classifier choice jointly affect discrimination, eleven mainstream machine-learning classifiers were trained on the feature outputs of each transfer-learning model: logistic regression (LR), Naïve Bayes, support vector machine (SVM), k-nearest neighbors (KNN), Random forest (RF), extremely randomized trees (ExtraTrees), XGBoost, LightGBM, gradient boosting (GB), AdaBoost, and multilayer perceptron (MLP). Core metrics—area under the curve (AUC), accuracy, sensitivity, and specificity—were computed in the test set as the base performance matrix.For each multichannel configuration, the corresponding set of 11 classifiers was analyzed as one group using summary receiver operating characteristic (SROC) methodology:

1. Use FPR = 1 − specificity and TPR = sensitivity as coordinates to fit a logistic-form SROC curve.

2. Compute the SROC area under the curve (SROC-AUC) to summarize group-level discriminative ability.

3. Calculate each model’s geometric distance to the ideal point (0, 1), and define the composite score:

   

   Where is the normalized AUC and is the normalized score derived from geometric distance.

4. Designate the model with the highest final_score within each group as the group-optimal model.

5. Repeat SROC fitting and scoring across all group-optimal models to identify, for each sequence (T1-w, CET1-w, T2-w), the global optimal deep transfer-learning model for subsequent fusion and comparative experiments.

To examine whether the selected sequence-specific configurations were sensitive to the weighting scheme used in the composite score, we performed a weighting-scheme sensitivity analysis. The weights assigned to the normalized AUC component and the SROC-distance component were varied from 0.9/0.1 to 0.1/0.9 in increments of 0.1. For each weighting scheme, the group-wise model-selection step and the final sequence-specific model-selection step were repeated for T1-w, CET1-w, and T2-w. Selection stability was assessed by the frequency with which the same sequence-channel-backbone configuration was selected across the tested weighting schemes.

For 2.5D model training, predefined training and validation files were used consistently for all candidate configurations. Image inputs were normalized using an ImageNet-style normalization strategy. Models were trained with a batch size of 32, stochastic gradient descent optimization, an initial learning rate of 0.01, and a maximum of 50 epochs. The same training protocol was applied to all candidate backbone and channel configurations to ensure fair comparison. Batch-level class balancing was not enabled in the current implementation.

Radiomics feature extraction

Radiomics features were extracted using the PyRadiomics package (https://pyradiomics.readthedocs) [24]. These features are not directly appreciable or quantifiable by visual inspection [25, 26]. In accordance with the Image Biomarker Standardisation Initiative (IBSI), a total of 107 radiomics features were computed from each region of interest (ROI) for each MRI sequence. The features comprised: (1) shape features (n = 14); (2) first-order statistics (n = 18); and(3) texture features, including gray-level co-occurrence matrix (GLCM, n = 24), gray-level run-length matrix (GLRLM, n = 16), gray-level size zone matrix (GLSZM, n = 16), gray-level dependence matrix (GLDM, n = 14), and neighboring gray-tone difference matrix (NGTDM, n = 5). Detailed definitions of these features are available in the cited literature [26].

Feature selection and single-modality model construction

Given that radiomics and deep-learning (DL) features may encode different structural and semantic information for survival prediction, feature selection and single-modality modeling were performed separately for radiomics features and for 2.5D multichannel DL features to ensure representative, non-redundant inputs. All features were first standardized using z-score normalization. Based on distributional characteristics, univariate screening was conducted with independent-samples t tests or Mann–Whitney U tests, retaining features that showed significant differences between the two classes (P < 0.05). Pairwise Pearson correlation coefficients were then computed, and the more informative or clinically interpretable variable was retained when high collinearity was present.To further reduce redundancy and improve generalization, least absolute shrinkage and selection operator (LASSO) regression was applied for feature selection; the regularization parameter λ was determined by five-fold cross-validation, and features with non-zero coefficients were retained as the final subset. Each final subset was then input to machine-learning classifiers for training, with hyperparameters optimized under a five-fold cross-validation scheme. Performance on the training and test sets was compared using area under the curve (AUC), accuracy, sensitivity, and specificity. For each feature family, the top-performing classifier was designated as the single-modality predictor for subsequent fusion and comparative analyses.

Multi-instance learning fusion

The fusion module was implemented as a MIL-inspired patient-level aggregation strategy. Each patient was defined as a bag, and the prediction probabilities generated from the sequence-specific radiomics and transfer-learning models were treated as instances within that bag. For the histogram representation, probabilities were grouped by patient identifier; for each selected probability column, a normalized histogram was calculated, and histograms from multiple columns were concatenated when applicable. For the BoW/TF-IDF representation, probability outputs were treated as patient-level corpus tokens, term frequency was calculated within each patient-level bag, inverse document frequency was calculated across patients, and the resulting TF-IDF features were used to emphasize discriminative probability-output patterns. The resulting patient-level representation was used for final EMFA risk estimation. This formulation aggregates heterogeneous model outputs at the patient level; it is not an attention-based MIL architecture and therefore does not directly generate attention-weight maps.

Statistical analysis

Continuous variables were analyzed using t tests or Mann–Whitney U tests according to distributional characteristics, and categorical variables were compared with the chi-square (χ²) test. In the test cohort, model performance was comprehensively evaluated by plotting receiver operating characteristic (ROC) curves and calculating multiple diagnostic indices, including the area under the ROC curve (AUC), specificity, sensitivity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) [27]. Agreement between nomogram-predicted and observed outcomes was assessed with calibration curves, and calibration was quantified using the Hosmer–Lemeshow goodness-of-fit test [28]. Clinical utility was examined via decision curve analysis (DCA). Analyses were performed using Python 3.7.12; part of the code execution and model analysis was conducted on the OnekeyAI platform (v3.5.12). Statistical procedures used statsmodels 0.13.2; radiomics feature extraction used PyRadiomics 3.0.1; machine-learning experiments used scikit-learn 1.0.2; and deep-learning models were implemented in PyTorch 1.11.0 with acceleration via CUDA 11.3.1 and cuDNN 8.2.1.

Results

Patient characteristics

In this retrospective study, 273 patients diagnosed with glioma were included and randomly assigned—using a stratified 7:3 split—to a training cohort (n = 191) and a test cohort (n = 82). Demographic and clinical characteristics are summarized in Table 1. The mean age was 59.69 ± 14.10 years in the training cohort and 57.48 ± 14.45 years in the test cohort (Table 1). Using the 24-month overall survival endpoint, the positive-to-negative class distribution was 49:142 in the training cohort and 18:64 in the test cohort.

Table 1.

Patient characteristics of the study sets

Variables	Training set	P value	Test set	P value	P value(Train vs. Test)
Age(years)	59.69 ± 14.10	< 0.001	57.48 ± 14.45	< 0.001	0.299
WHO grades		0.074		< 0.001	0.458
2	4(2.09%)		4(4.88%)
3	101(52.88)		42(51.22%)
4	86(45.03%)		36(43.90%)

Feature extraction and selection

Experiments were conducted using radiomics and deep transfer-learning (DTL) features derived from three MRI sequences: (T1-w, CET1-w, T2-w).

1. Radiomics features

   A total of 321 radiomics features were extracted, each sequence was 107 radiomics features.

2. DTL features

   For T1-w, CET1-w, and T2-w, ResNet-34, ResNet-18, and ResNet-34 backbone models, respectively, were trained. From each model, 512 DTL features were extracted and then reduced to 256 components via principal component analysis (PCA) to lower dimensionality and mitigate multicollinearity.

3. Feature selection

   On the PCA-reduced feature sets, least absolute shrinkage and selection operator (LASSO) regression was applied, yielding 46 features for downstream modeling: 19 DTL features (7 from T1-w, 7 from CET1-w, 5 from T2-w) and 27 radiomics features (12 from T1-w, 10 from CET1-w, 5 from T2-w). Detailed results of feature extraction/selection and the corresponding equations used for feature-importance scoring are provided in Supplementary File 1.D.

Evaluation of channel configurations and architectures

To systematically assess the impact of channel configuration and backbone depth on predictive performance, multichannel transfer-learning models were constructed for each MRI sequence. Models were uniformly named [Sequence]_[Channel]C_[ResNet]X, Sequence = T1-w, CET1-w, T2-w; Channel = 3,5,7,9; X = 18,34,50,101,152. At evaluation, feat ure outputs fromeach model were used to train 11 conventional machine-learning classifiers (LR, NaiveBayes, SVM, KNN, RandomForest, ExtraTrees, XGBoost, LightGBM, GradientBoosting, AdaBoost, MLP), and AUC, accuracy, sensitivity, and specificity were computed. T he 11 classifier results associated with a given backbone constituted one group. For each group—and for the classifiers within groups—summary ROC (SROC) curves were fitted with a logistic function to obtain the group-level mean SROC-AUC and the geometric distance to the ideal point (0, 1). A composite score was then calculated to quantify group-level discriminative ability. As shown in Fig. 3a–c, the groups T1_9C_ResNet34, T1CE_3C_ResNet18, and T2_9C_ResNet34 achieved the highest composite scores within their respective sequences, with SROC-AUCs of 0.829, 0.871, and 0.879; the corresponding SROC areas are depicted in Fig. 3d–f. Within-group comparisons further indicated that LightGBM performed best in the T1_9C_ResNet34 group (Fig. 4d), SVM in the T1CE_3C_ResNet18 group (Fig. 4e), and LightGBM in the T2_9C_ResNet34 group (Fig. 4f). Accordingly, T1_9C_ResNet34, T1CE_3C_ResNet18, and T2_9C_ResNet34 were designated the final sequence-wise optimal models and were carried forward for subsequent fusion and comparative experiments. Complete results for all transfer-learning models under T1-w, CET1-w, and T2-w are provided in Supplementary Files 2–4.

Fig. 3.

SROC curve fitting results for multichannel transfer-learning models across three MRI sequences. (a–c) Strict SROC fits for all model groups under the T1-w, CET1-w, and T2-w sequences, respectively. (d–f) SROC fits and corresponding areas (SROC-AUC) for the group with the highest composite score within each sequence. The optimal groups were T1_9C_ResNet34 (SROC-AUC = 0.829) for T1, T1CE_3C_ResNet18 (SROC-AUC = 0.871) for T1CE, and T2_9C_ResNet34 (SROC-AUC = 0.879) for T2-w. Green markers indicate the best classifier within each optimal group: LightGBM (T1), SVM (CET1-w), and LightGBM (T2-w)

Fig. 4.

ROC-curve comparisons across models in the training and test cohorts. (a–c) ROC curves of the best-performing classifiers within the radiomics models for the T1-w, CET1-w, and T2-w sequences, shown for both training and test cohorts.(d–f) ROC curves of the best-performing classifiers within the sequence-wise optimal transfer-learning (TL) models for T1-w, CET1-w, and T2-w, shown for both training and test cohorts

The weighting-scheme sensitivity analysis demonstrated stable sequence-specific model selection across alternative composite-score weights. Across the nine tested alpha/beta settings, the originally selected configurations were consistently retained for all three sequences: T1_9C_ResNet34 for T1-w, T1CE_3C_ResNet18 for CET1-w, and T2_9C_ResNet34 for T2-w. Each configuration was selected in 9/9 weighting settings, indicating that the final sequence-channel-backbone choices were robust to the relative weighting of the AUC and SROC-distance components. The detailed results are provided in Supplementary File 1.

Performance comparison of the EMFA framework

To validate the proposed EMFA framework, head-to-head comparisons were conducted against (i) the sequence-wise optimal single-sequence models for T1-w, CET1-w, and T2-w, and (ii) representative early-fusion baselines. Figure 5a displays the ROC curves of the optimal radiomics classifiers for the T1-w, CET1-w, and T2-w models, respectively, and Fig. 5b shows the ROC curves of the sequence-wise optimal transfer-learning classifiers for the same three sequences. Figure 5c summarizes the cross-model comparisons. EMFA achieved strong discriminative performance in both the training and test cohorts. In the test cohort, EMFA reached an AUC of 0.899 (95% CI, 0.711-1.000), with accuracy of 0.935 (93.5%) and sensitivity of 88.9%, indicating competitive performance relative to the single-sequence and early-fusion models. These results suggest that multisequence, multichannel fusion may help harness complementary information and enhance discrimination. Table 2 summarizes the primary metrics in the training and test cohorts; however, the comparative results in the held-out test cohort should be interpreted together with the DeLong tests and confidence intervals reported below. To further characterize potential data-source heterogeneity, we performed a source-stratified analysis within the held-out test cohort. EMFA achieved an AUC of 0.899 and accuracy of 0.935 in the overall test cohort. In the private institutional subset, the AUC, accuracy, sensitivity, specificity, PPV, and NPV were 0.821, 0.865, 0.803, 0.861, 0.625, and 0.824, respectively. In the BraTS-derived subset, the corresponding values were 0.745, 0.787, 0.778, 0.704, 0.608, and 0.861. These source-stratified results suggest that the model retained discriminative ability in both subsets, but performance was lower than the pooled test estimate, indicating possible data-source heterogeneity and domain shift.

Fig. 5.

ROC-curve comparisons across models in the training and test cohorts. (a) ROC comparison between the EMFA model and the three single-sequence radiomics models (T1-w, CET1-w, T2-w) in the test cohort. (b) ROC comparison between the EMFA model and the three sequence-wise optimal TL models in the test cohort. (c) ROC comparison between the EMFA model and three early-fusion baselines (radiomics early fusion, TL early fusion, and all early fusion) in the test cohort. Complete classifier-level performance for single-sequence radiomics and TL models is provided in Supplementary File 1.E, and full ROC curves comparing EMFA with all comparators are provided in Supplementary File 1.F

Table 2.

Diagnostic performance of the EMFA framework in the training and test cohorts

Cohort	Model	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV
Train	EMFA	0.958	0.993	0.959	0.957	0.887	0.985
Train	T1	0.826	0.859	0.878	0.809	0.614	0.950
Train	T1CE	0.763	0.880	0.878	0.723	0.524	0.944
Train	T2	0.832	0.902	0.878	0.816	0.623	0.950
Train	T1_9C_ResNet34	0.885	0.947	0.939	0.866	0.706	0.976
Train	T1CE_3C_ResNet18	0.932	0.946	0.898	0.944	0.846	0.964
Train	T2_9C_ResNet34	0.812	0.902	0.878	0.789	0.589	0.949
Train	Radiomics early fusion	0.791	0.833	0.837	0.775	0.562	0.932
Train	TL early fusion	0.880	0.906	0.714	0.937	0.795	0.905
Train	All early fusion	0.885	0.946	0.857	0.894	0.737	0.948
Test	EMFA	0.935	0.899	0.889	0.955	0.889	0.955
Test	T1	0.935	0.859	0.778	1.000	1.000	0.917
Test	T1CE	0.871	0.919	0.889	0.864	0.727	0.950
Test	T2	0.839	0.871	0.778	0.864	0.700	0.905
Test	T1_9C_ResNet34	0.774	0.823	0.778	0.773	0.583	0.985
Test	T1CE_3C_ResNet18	0.871	0.894	0.889	0.864	0.727	0.950
Test	T2_9C_ResNet34	0.935	0.879	0.778	1.000	1.000	0.917
Test	Radiomics early fusion	0.806	0.833	0.778	0.818	0.636	0.900
Test	TL early fusion	0.871	0.939	0.889	0.864	0.727	0.950
Test	All early fusion	0.903	0.896	0.778	0.955	0.875	0.913

To assess statistical differences in discriminative performance, we performed paired DeLong tests to compare the AUCs of the EMFA framework with those of each single-sequence model and each early-fusion baseline (Fig. 6a–f). In the training cohort (Fig. 6a, c, e), EMFA achieved significantly higher AUCs than all three single-sequence radiomics models (T1-w, CET1-w, and T2-w; all p < 0.001; Fig. 6a). Consistent superiority was also observed when EMFA was compared with the sequence-wise optimal transfer-learning (TL) models (T1_9C_ResNet34, T1CE_3C_ResNet18, and T2_9C_ResNet34; p = 0.006, 0.029, and < 0.001, respectively; Fig. 6c). Moreover, EMFA significantly outperformed the early-fusion baselines (radiomics early fusion, TL early fusion, and all early fusion; p < 0.001, 0.001, and 0.001, respectively; Fig. 6e), supporting the effectiveness of the proposed staged fusion strategy.

Fig. 6.

DeLong tests for AUC differences between models. (a, b) EMFA versus the radiomics models for the T1-w, CET1-w, and T2-w sequences in the training (a) and test (b) cohorts. (c, d) EMFA versus the sequence-wise optimal transfer-learning models in the training (c) and test (d) cohorts. (e, f) EMFA versus three early-fusion baselines (radiomics early fusion, TL early fusion, and all early fusion) in the training (e) and test (f) cohorts

In the test cohort (Fig. 6b, d, f), none of the pairwise contrasts reached statistical significance (all p > 0.05), indicating that EMFA maintained competitive discriminative performance relative to the single-sequence and early-fusion baselines rather than statistically significant superiority in the held-out test cohort. Taken together, these results demonstrate clear gains for EMFA during training and stable but statistically uncertain comparative performance on the held-out test set; therefore, the test-cohort findings should be interpreted cautiously in light of the limited sample size and wide confidence intervals.

To enhance clinical interpretability, individualized risk scores from the EMFA model were integrated with available clinical variables, including age and WHO grade, to construct an integrated nomogram (Fig. 7). This nomogram should be interpreted as a clinical-imaging risk presentation based on the variables available in this retrospective cohort rather than as a complete contemporary molecular-clinical prognostic model. To further assess predictive reliability and clinical utility in the test cohort, we performed the Hosmer–Lemeshow goodness-of-fit test and calibration curve analysis. As shown in Fig. 8a–c, the EMFA model achieved a p-value of 0.804 and a Brier score of 0.189, indicating good agreement between predicted probabilities and observed outcomes; comparator calibration curves deviated more from the ideal line, suggesting less stable probabilistic predictions in this cohort. Clinical utility, evaluated by decision curve analysis (DCA), is presented in Fig. 8d–f. Across threshold probabilities, EMFA showed favorable net benefit relative to the single-sequence and early-fusion models. Collectively, these results suggest that EMFA offers competitive discriminative performance, predictive reliability, and clinical interpretability.

Fig. 7.

Integrated nomogram combining the EMFA risk score with available clinical variables. The revised nomogram presentation includes clearer variable labels, scale markings, and risk-score annotation to support individualized risk interpretation

Fig. 8.

Calibration curves and decision curve analysis (DCA) in the test cohort. (a) Calibration curves for the EMFA model versus the single-sequence radiomics models (T1-w, CET1-w, T2-w). (b) Calibration curves for EMFA versus the sequence-wise optimal transfer-learning models. (c) Calibration curves for EMFA versus the three early-fusion models. (d) DCA curves for EMFA versus the single-sequence radiomics models (T1-w, CET1-w, T2-w). (e) DCA curves for EMFA versus the sequence-wise optimal transfer-learning models. (f) DCA curves for EMFA versus the three early-fusion models. Full calibration and DCA comparisons between EMFA and all models are provided in Supplementary File 1.G

Discussion

The study proposed an EMFA framework that integrates 2.5D multi-channel inputs from multisequence MRI, transfer-learning features, and radiomics, while employing a MIL-inspired patient-level aggregation strategy to integrate heterogeneous probability outputs across sequences and feature types. From a feature-level perspective, relying solely on either radiomics or deep learning features presents inherent limitations. Deep learning possesses the capacity to automatically extract nonlinear representational patterns but lacks interpretability [29], whereas radiomic features are structurally explicit and interpretable but are easily influenced by subjective segmentation and handcrafted assumptions [30]. On a held-out test set, the model achieved an AUC of 0.899 and an accuracy of 0.935 (93.5%). However, pairwise DeLong comparisons in the test cohort did not reach statistical significance; therefore, these results are interpreted as indicating competitive and clinically useful performance rather than definitive statistical superiority. Calibration and clinical utility were favorable, with a Hosmer-Lemeshow test p value of 0.804, a low Brier score among comparators, and decision-curve analysis demonstrating useful net benefit across a range of risk thresholds. Coupled with a risk score and a nomogram, the framework yields an intuitive decision-support tool for individualized survival risk assessment.

Multisequence MRI provides inherently complementary lesion characterization: the T1-w sequence provide basic tissue structure and can be used to identify morphological features of tumors [31], the contrast-enhanced T1-weighted (CET1-w) sequence is sensitive to contrast enhancement and blood–brain barrier disruption, and the T2-w sequence better depicts infiltration and edema boundaries [32]. Prior evidence indicates that cross-sequence integration can improve discriminative stability—for example, Niu et al. combined FLAIR and CET1-w to predict IDH-wildtype plus TERT-mutant status [33]. Building on this rationale, the studyjointly leveraged T1-w, CET1-w, and T2-w sequences to achieve more comprehensive lesion representation and better alignment with clinical scenarios [32]. From a methodological perspective, the study adopted a multi-channel 2.5D design paired with ResNets of varying depth—an approach repeatedly validated across diverse tasks. He et al. employed a seven-channel CT model to predict T1N0 gastric cancer [34], and Sun et al. applied a similar framework for lymphovascular invasion prediction [35]. In MRI-based studies, Liu et al. utilized a Transformer-based 2.5D model for parotid tumor classification [36]; Li et al. used ResNet-50 for glioma subtype stratification and prostate cancer aggressiveness [37], respectively; Zhang et al. applied ResNet-50 for hepatocellular carcinoma recurrence prediction [38]; and Yin et al. confirmed the advantages of 2.5D multi-channel architectures in lung adenocarcinoma [39]. To obtain a reproducible and quantitatively optimal configuration, the study systematically evaluated combinations of ResNet-18/34/50/101/152 with 3/5/7/9 channels and employed summary ROC (SROC) analysis to derive a unified “sequence–channel matching” strategy. The results indicated a preference for nine channels for T1-w and T2-w, and three channels for CET1-w, underscoring that a single channel depth cannot simultaneously accommodate the distinct spatial characteristics of each sequence. Notably, some [Sequence]_[Channel]C[ResNet]X combinations in Supplementary Files 2-4 did not yield effective features during preliminary screening and were therefore excluded from the main results. These combinations exhibited near-random feature importance, LASSO solutions collapsing to near-zero or instability, cross-validation AUCs around 0.5 with large variance, and no incremental net benefit on calibration or decision curves. The likely causes include over-parameterization and training instability in a small-sample regime (high memory footprint, gradient explosion/vanishing, early convergence followed by generalization drop), as well as mismatches between intrinsic sequence-specific texture frequencies and the network’s receptive-field or stride settings. Guided by robustness and clinical usability, only configurations that passed consistency checks and multi-metric validation were reported. Building upon this framework, the introduction of the MIL mechanism yielded notable comprehensive benefits: (1) At the cross-sequence level, dynamic attention enables a “soft-selection” mechanism that adaptively focuses on key pathological patterns across sequences (e.g., enhancement ring and necrotic core on CET1-w, and hyperintense infiltrative rims on T2-w/FLAIR); (2) At the feature-aggregation level, MIL promotes alignment across sequence-specific feature distributions and strengthens the complementary representation of multimodal information; and (3) For interpretability, MIL-derived attention heatmaps improve concordance between model focus and radiologic priors, jointly enhancing discriminative power, stability, and explainability.

Compared with prior glioma survival-prediction studies, EMFA occupies an intermediate position between handcrafted radiomics models and fully volumetric deep-learning approaches. Radiomics-based methods provide interpretable quantitative descriptors but may be sensitive to segmentation and feature-engineering choices, whereas 3D CNN and Transformer-based approaches can capture richer volumetric context but generally require larger datasets and greater computational resources. Graph-based and MIL-related methods have also been explored to model regional heterogeneity and patient-level aggregation. Direct numerical comparison across these studies is limited by differences in patient inclusion, survival endpoint definition, censoring treatment, data partitioning, molecular-marker availability, and reported metrics. Within this context, the main methodological rationale of EMFA is to use 2.5D multichannel inputs to balance spatial context and computational efficiency, while integrating multisequence MRI, transfer-learning features, and radiomics features to provide complementary patient-level representations.

FLAIR was not included in the main framework because the present retrospective analysis required a consistent sequence set across the institutional historical cohort and the BraTS-derived cases. Although FLAIR is clinically valuable for visualizing infiltrative tumor margins and peritumoral edema, its availability and protocol consistency were limited in the institutional historical data. We therefore focused on T1-w, CET1-w, and T2-w, which were more consistently available across both data sources. The absence of FLAIR remains an important limitation, and future work should evaluate whether FLAIR provides incremental prognostic value in cohorts with complete and standardized multisequence MRI.

The study has several limitations: The retrospective design and mixed data sources constrain generalizability across institutions and scanners; incomplete long-term follow-up in certain patients limits robust evaluation of distant outcomes; and IDH mutation status, MGMT promoter methylation, and TERT promoter mutation were not consistently available across the institutional historical cohort and BraTS 2021-derived cases. Because these molecular markers are major determinants of glioma subtype and prognosis, the present imaging-based framework cannot quantify incremental predictive value beyond a contemporary clinical-molecular baseline model. The current framework also does not incorporate cross-modal information such as DWI/ADC, DSC/DCE perfusion, DTI, MRS, or pathology/genomics beyond the available clinical variables, leaving potentially informative biological features underexplored. Future work will focus on multi-center prospective cohorts with complete molecular-pathologic annotation, direct comparison among clinical-molecular baseline models, imaging models, and combined models, batch-effect correction and domain adaptation, lightweight 3D and Transformer architectures, and multimodal fusion integrating imaging, pathology, and genomics to enhance spatial modeling capacity, cross-domain robustness, and clinical deployability. Although the SROC-based selection strategy is not a conventional external validation procedure and the evaluation of multiple model configurations may introduce selection bias, the weighting-scheme sensitivity analysis showed that the selected configurations were stable across alternative composite-score weights. This finding reduces concern that the final sequence-channel-backbone choices were artifacts of a single arbitrary weighting scheme; nevertheless, independent external validation remains necessary. In addition, the present aggregation module is MIL-inspired but not attention-based; therefore, attention-weight maps are not directly available. Future work should incorporate explicit attention-based MIL, saliency mapping, or SHAP-based analyses to provide more complete interpretability. Although the combined cohort was divided into training and held-out test sets, this random split does not constitute true external validation because institutional and BraTS-derived cases were pooled before splitting. Additional independent multi-center validation, including source-specific testing or training on one data source and testing on another, is needed to determine cross-institution and cross-scanner generalizability. Detailed treatment protocols were incomplete and not harmonized across the retrospective institutional cases and BraTS-derived cases; therefore, treatment factors such as extent of resection, radiotherapy, chemotherapy, and specific regimens could not be incorporated into model training or adjustment. This missing treatment information may introduce bias, and the current EMFA framework should be interpreted as a preoperative imaging-based risk model rather than a treatment-adjusted prognostic model. The relatively high training AUC compared with the test AUC suggests a potential risk of overfitting. In the current 2.5D model-training implementation, explicit weight decay, early-stopping patience, and systematic data augmentation were not specified, which may limit regularization strength. Future studies should incorporate predefined regularization strategies, data augmentation, early stopping, nested validation, and larger external cohorts to improve generalizability. The source-stratified analysis showed lower performance in the private institutional subset and BraTS-derived subset than in the pooled test cohort, supporting the possibility of data-source heterogeneity and domain shift. Therefore, the current held-out test results should be interpreted as retrospective internal evidence rather than definitive external validation. In addition, the final ROIs were consensus-reviewed masks generated through primary delineation and secondary expert review. This workflow is a practical quality-control strategy for retrospective imaging studies and can reduce obvious contouring errors before feature extraction. However, it cannot replace a formal inter-reader or intra-reader reproducibility analysis. Because independent preliminary masks were not systematically archived for all retrospective cases, Dice similarity coefficients or ICC values for segmentation reproducibility could not be reliably calculated retrospectively. Future prospective studies should preserve independent reader segmentations and report Dice/ICC-based reliability analyses, particularly for radiomics features that may be sensitive to boundary variation.

In conclusion, the EMFA framework—anchored by “2.5D multi-channel inputs + SROC-guided sequence–channel matching + complementary transfer-learning × radiomics features + MIL aggregation” — demonstrated competitive discrimination, calibration consistency, and clinical net benefit in this retrospective cohort. Through the clarified patient-level aggregation workflow and a nomogram interface, the framework is both methodologically reproducible and engineering-ready, providing a scalable pathway for preoperative survival prediction and individualized clinical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

PCL and YMR were responsible for the study design, protocol development, literature review, model design and construction, and manuscript writing. YYQL and MW contributed to the data collection and supervised the data analysis. YWM participated in data collection and critically revised the manuscript’s important content. JRC, YWM, and XYH oversaw the execution of research activities. All the authors reviewed and approved the final manuscript.

Funding

Not applicable.

Data availability

The private datasets used and analysed during this study are available from the corresponding author upon reasonable request and subject to institutional ethics approval and data-use restrictions. The BraTS-derived data are available through the original public challenge/data-access mechanism. Public-release analysis scripts with local paths removed are available at GitHub: https://github.com/Newbegin23/EMFA.git. Because part of the analysis was conducted on the OnekeyAI platform and some files are restricted by data-use agreements, additional code, processed non-sensitive materials, and implementation details may be made available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki and was approved by the institutional ethics review board of the First Affiliated Hospital of Dalian Medical University. The requirement for informed consent was waived by the ethics committee in accordance with national regulations for retrospective studies involving anonymized data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.