Exploring Deep Learning and Hybrid Approaches in Molecular Subgrouping and Prognostic Related Genetic Signatures of Medulloblastoma

Exploring Deep Learning and Hybrid Approaches in Molecular Subgrouping and Prognostic Related Genetic Signatures of Medulloblastoma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Exploring Deep Learning and Hybrid Approaches in Molecular Subgrouping and Prognostic Related Genetic Signatures of Medulloblastoma Yanong Li, Hailong Liu, Yawei Liu, Jane Li, Hiro Hiromichi Suzuki, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6622165/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 15 Sep, 2025 Read the published version in Chinese Neurosurgical Journal → Version 1 posted You are reading this latest preprint version Abstract Objectives Deep learning (DL) based on MRI of medulloblastoma enables risk stratification, potentially aiding in therapeutic decisions. This study to develop DL models that identify four medulloblastoma molecular subgroups and prognostic related genetic signatures. Materials and Methods In this retrospective study, consecutive patients with newly diagnosed MB at MRI (T1-, T2- and contrast-enhanced T1-weighted) at two medical institutes between January 2015 and June 2023 were identified. Two-stage sequential DL models were designed—MB-CNN that first identifies wingless (WNT), sonic hedgehog (SHH), Group 3, and Group 4. Further, prognostic related genetic signatures DL models (MB-CNN_TP53/MYC/Chr11) were developed to predict TP53 mutation, MYC amplification, and chromosome 11 loss status. A hybrid model combining MB-CNN and conventional data (clinical information and MRI features) was compared to a logistic regression model constructed only with conventional data. Four-classification tasks were evaluated with confusion matrices (accuracy) and two-classification tasks with ROC curves (area under the curve, AUC). Results The datasets comprised 449 patients (mean age ± SD at diagnosis, 13.55 years ± 2.33, 249 males). MB-CNN accurately classified MB subgroups in the external test dataset (accuracy was in the range of 76.29–78.71%). MB-CNN_TP53/MYC/Chr11 models effectively predicted signatures (AUC of TP53 in SHH: 0.91, MYC amplification in Group 3: 0.87, chromosome 11 loss in Group 4: 0.89). The accuracy of hybrid model outperformed the logistic regression model (82.20% vs. 59.14%, P = .009) and showed comparable performance to MB-CNN (82.20% vs. 77.50%, P = .105). Conclusion MRI-based DL models allowed identification of the molecular medulloblastoma subgroups and prognostic related genetic signatures. Medulloblastoma subgroups MRI Deep learning Risk stratification Convolutional neural network Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Medulloblastoma (MB), the most prevalent malignant pediatric brain tumor in the posterior fossa( 1 ), is molecularly classified into four subgroups under the 2021 WHO framework: Wingless (WNT), Sonic Hedgehog (SHH), Group 3 (G3), and Group 4 (G4)1–2( 2 ). Prognosis varies significantly across subgroups, with WNT tumors demonstrating > 90% five-year survival, SHH subgroups exhibiting TP53 mutation-dependent heterogeneity, and MYC-amplified G3/G4 cases showing aggressive behavior and poor therapeutic response. This molecular stratification underpins current risk-adapted clinical management strategies( 3 – 11 ). Deep learning (DL) has revolutionized preoperative molecular prediction in gliomas, enabling accurate identification of biomarkers such as MGMT promoter methylation and IDH mutations. While preliminary DL applications in MB achieved 85% subgroup classification accuracy( 12 , 13 ). Critical gaps persist in current MB prognostic models, including insufficient cohort sizes for robust generalization, exclusion of high-risk genetic signatures (TP53 mutations, MYC amplification, chromosome 11 loss) in risk stratification frameworks, and underutilized opportunities to enhance predictive accuracy through multimodal clinical-MRI data integration( 14 , 15 ). To address these limitations, we develop a dual DL framework integrating preoperative MRI with methylation/NGS data. First, MB-CNN classifies tumors into molecular subgroups. Second, hybrid models (MB-CNN_TP53, MB-CNN_MYC, MB-CNN_Chr11) predict high-risk genetic signatures specific to each subgroup. Finally, a clinical-MRI fusion model combines radiomic features, conventional imaging biomarkers, and clinical variables to optimize prognostic accuracy. This approach bridges the translational gap between molecular subgroup classification and actionable genetic risk assessment. Materials and Methods Study design and population This study comprised two sequential stages. The initial stage developed a predictive DL model (MB-CNN) for molecular subgroups of MB (WNT, SHH, G3, and G4). The second stage constructed further DL models to predict prognostic related genetic signatures, TP53 mutation (MB-CNN_TP53), MYC amplification (MB-CNN_MYC), and chromosome 11 loss (MB-CNN_Chr11), within the SHH, G3, and G4 subgroups. Furthermore, additional analyses were performed. Figure 1 shows the overall study flowchart. A total of 449 patients with MB in two independent medical institutes were consecutively enrolled (shown in Method S1 ). This multicenter retrospective study was approved by the institutional review board (Ethics committee approval No. 82172608), with consent waived due to minimal risk. The inclusion criteria were as follows: (i) All patients diagnosed with MB confirmed between January 2015 and June 2023; (ii) necessary preoperative MR scans including axial T2-weighted (T2W), T1-weighted (T1W), and contrast-enhanced T1-weighted (CE-T1W) sequences. The exclusion criteria were as follows: (i) patients who lack the molecular subgroup and prognostic related genetic data; (ii) patients with prior neurosurgical procedures before MR scans; (iii) MR images with severe artifacts affecting radiological assessment. Imaging preprocessing Brain MR images were performed with at 1.5- or 3.0-T magnet (Philips Healthcare, Siemens Healthineers, GE Healthcare, and Toshiba Medical Systems USA). The MRI scanning details are in Method S2 . Figure S1 shows the proportion of manufactures across datasets. Preprocessing initiated with conversion to the NIfTI format (FSLv6.0, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/ ) for data consistency, followed by spatial resampling and intensity normalization within a range of -1 to 1 to reduce variability across scanners and improve comparability. Method S5 contains comprehensive explanations of these methodological refinements to ensure data consistency and dependability across MR systems. Blinded tumor segmentation for ground truth generation We, integrated methylation arrays, next-generation sequencing technologies (Whole Genome Sequencing (WGS) and Comparative Genomic Hybridization (CGH)), to accurately identify four molecular subgroups (presented in Method S4 )( 16 , 17 ). For medical image segmentation, the Unet architecture was selected for its proven effectiveness( 18 ). Its Encoder-Decoder design extracts high-dimensional features via convolutional and max pooling layers, while up-convolution layers reconstruct segmentation maps for precise tumor identification. The segmentation model was developed in the development dataset of 325 patients’ MR scans (T1W, CE-T1W, T2W), split into 80% training (n = 260), 10% validation (n = 33), and 10% testing (n = 32). Training of the Unet model was conducted on the PyTorch ( https://pytorch.org/ ) over 100 epochs using the Adam optimizer with a cosine decay learning rate schedule. The Dice Loss function serves as the performance metric, with the model’s effectiveness periodically assessed via the Dice Score on the validation dataset. To establish a robust ground truth for the DL model, two neuroradiologists (Y.N.L. and D.C.., > 6 years experience) independently reviewed the segment results of Unet model. Low-quality images were resegmented using ITK-SNAP (version 3.8.0, http://www.itksnap.org ) and refined by a senior neuroradiologist (Y.O.L., 20 years experience). Overlapping agreed regions served as the final ground truth for model development. Conventional MR characteristics evaluation Two experienced neuroradiologists (Y.N.L. and D.C.) independently evaluated conventional MR images characteristics across all datasets, including T1W foci-hyperintensity, T2W foci-hypointensity, and levels of enhancement on CE-T1W images (presented in Method S3 ). The agreement between readers was excellent (Kappa scores: 0.842–0.864, P < .001). Discrepancies were adjudicated by a senior neuroradiologist (Y.O.L.). DL models training and validation In this study, we developed MB-CNN and MB-CNN_TP53/MYC/Chr11 based on preoperative MR images (T1W, CE-T1W, and T2W). We modified the state-of-the-art 3D nnU-Net framework( 19 ) to segment the tumor and predict molecular subgroups and genetic signatures. The ResNet-50 architecture, characterized by its advanced deep residual learning strategy, incorporates a 50-layer deep structure designed to overcome the vanishing gradient issue common in extensive neural networks. Its residual blocks enhance gradient propagation throughout the network, enabling the efficient training of deeper neural models. During the initial stage, for the MB-CNN project, ResNet-50 classified tumor molecular subgroups. Our study leveraged a dataset encompassing MR images from 325 patients, identical to the cohort used for segmentation model training. This dataset was methodically partitioned into training and validation datasets following 3:1 ratio (244:81), ensuring a randomized distribution. Additionally, a set of 124 independent external patients was designated as the test dataset. Each dataset entry encompasses a tumor segmentation map derived from the original MR scans, facilitated by the corresponding Unet model. The strategic dataset segmentation ensured equitable representation across the four molecular subgroups. When the MB-CNN predicts the direct output result of WNT, and when the model predicts SHH, G3 or G4 subgroups, the prediction process will enter the second stage-three specialized two-class DL models (MB-CNN_TP53, MB-CNN_MYC, and MB-CNN_Chr11), based on ResNet-50 to assess prognostic genetic signatures. The models processed raw MR images and tumor segmentation maps produced by Unet. Prior to final classification through a fully connected layer, the images were resized to dimensions of 256x256x3 and underwent convolutional processing within the ResNet-50 framework. Subsequently, our model training utilized prognostic related genetic signatures status from SHH (n = 54), G3 (n = 78), and G4 (n = 83) patient diagnoses. For validation, the model employed prognostic related genetic signatures identified within the validation dataset (SHH: n = 14, G3: n = 20, G4: n = 21), with independent external test data serving for comprehensive testing (SHH: n = 44, G3: n = 23, G4: n = 38). The classification model underwent 10 intensive training cycles on the PyTorch ( https://pytorch.org/ ) to enhance its accuracy in classification and prediction tasks. PyTorch was chosen for its support of dynamic computational graphs and GPU acceleration, essential for handling the ResNet-50 architecture and large datasets. To address class imbalance, the FocalLoss function prioritized hard-to-classify instances, enhancing sensitivity to underrepresented classes. The Adam optimizer adjusted learning rates for efficient convergence, refined by a cosine decay schedule. To evaluate the model’s generalizability and robustness, an independent dataset of 124 patients was tested, providing insights into the model's real-world applicability. For four-class classification, the MB-CNN model outputs a four-element vector of probabilities, each representing the likelihood of the input belonging to a specific category, ensured by the SoftMax activation function ( Eq. 1 ). \(\:Softmax\left({Z}_{i}\right)=\frac{{e}^{{Z}_{i}}}{{\sum\:}_{j}\:{e}^{{Z}_{j}}}\) ( Eq. 1 ) Here, \(\:{\text{Z}}_{\text{i}}\) is the value of the \(\:\text{i}\) th element in vector \(\:\text{Z}\) , and the denominator is the sum of applying the \(\:{\text{e}}^{\text{x}}\) function to all elements. Here, \(\:\text{e}\) is the base of the natural logarithm, approximately equal to 2.71828. In binary classification tasks, the model simplifies its output to a single probability value, indicating the likelihood of the input belonging to the "positive" class, a method that streamlines the output for binary tasks while maintaining predictive reliability. The DL model segmentation and classification architectures are shown in Fig. 2 A. Further, we independently tested it in the independent external test dataset. The Method S6 and S7 provide the training details of the segmentation and classification models. DL model evaluation The predictive accuracy of the MB-CNN was rigorously assessed using a confusion matrix against the ground truth. The model produced probability scores for each subgroup, facilitating a probabilistic interpretation of results. For classification purposes, each instance was assigned to the molecular subgroup corresponding to the highest probability score. Accuracy and recall for each class were derived from the confusion matrix, offering insights into the model's performance on a per-class basis. These metrics were complemented by precision and the F1 score, which are particularly informative in scenarios of class imbalance, providing a balanced view of the model's predictive performance. The predictive accuracy of the second stage DL model (MB-CNN_TP53/MYC/Chr11) was assessed against ground truth via receiver operating characteristic (ROC) analysis. The model generated continuous probability scores for each subgroup. The optimal cutoff value for classifying subgroups was determined in the training dataset based on the Youden index, which maximizes the combined sensitivity and specificity. Subsequently, the molecular subgroup designation was assigned based on the highest predicted probability. The model's performance was evaluated using standard metrics, including the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Further, to ensure a robust evaluation, an independent external dataset comprising 105 patient samples (SHH: n = 44, G3: n = 23, G4: n = 38) was utilized for testing. This external test step is critical for assessing the model's generalization capability and reliability in real-world scenarios, extending beyond the initial training and validation datasets. Additional analyses for clinical applicability To further elucidate the potential application of the DL model in diverse clinical scenarios, we conducted two complementary sets of additional analyses. First, we constructed a logistic regression model using clinical information (age and gender) and conventional MR characteristics as a baseline for comparison. Subsequently, we built a hybrid model using logistic regression model incorporating the outputs of the MB-CNN. Ultimately, these analyses aim to evaluate the standalone performance of the MB-CNN. Additionally, they seek to assess the potential enhancive value provided by integrating ancillary inputs, such as clinical data and MR characteristics, into the MB-CNN framework. Finally, we divided the independent test dataset into adult group (> 18 years, n = 26) and juvenile group (≤ 18 years, n = 98) according to age to compare MB-CNN and MB-CNN_TP53/MYC/Chr11 The models' performance was evaluated using confusion matrix (accuracy, recall, precision, and F1 score) and ROC curve (AUC, accuracy, sensitivity, and specificity). Statistical analysis We used the Statistical Package for the Social Sciences (SPSS) software (version 22, IBM, USA) and Python (version 3.6, http://www.python.org ) for statistical analyses. Categorical variables were displayed as frequency and percentages and tested using Pearson's Chi-squared test or Fisher's exact test. Continuous variables were displayed as mean and standard deviation (SD) and were tested using a two-sample t-test or Mann–Whitney U test (the choice between the t-test and Mann–Whitney U test was based on normality tests). A two-sided P < .05 was considered statistically significant. Method S8 shows more details of the Statistical analysis. Code Availability Codes are available via: https://github.com/YanongL/DL-model-code-for-MB-subgroups Results Patient characteristics Our study cohort consecutive included 449 patients diagnosed with MB, comprising 249 males and 200 females [mean age = 13.63 ± 2.47]. The age distribution revealed three distinct groups: 14 patients younger than 3 years [mean age = 1.86 ± 1.06], 243 patients aged 3–18 years [mean age = 11.30 ± 3.12], and 68 adults aged 18–49 years [mean age = 28.67 ± 4.33]. Detailed demographic information, genetic profiling, and MR characteristics of the study population are provided in Table 1 . Table 1 The main demographic and conventional MR features of the training dataset, validation dataset, and the independent external test dataset. Development Set Independent External Test Dataset Training dataset (n = 244) Validation dataset (n = 81) (n = 124) Molecular subgroups WNT SHH G3 G4 WNT SHH G3 G4 WNT SHH G3 G4 Count, n 45 76 78 45 20 24 28 9 21 19 45 39 Demographic Age (mean ± SD), y 14.42 ± 1.33 13.11 ± 2.47 12.53 ± 3.21 Gender, n Female: Male 22:23 40:36 29:49 12:33 8:12 15:9 8:20 2:7 9:12 10:19 28:17 17:12 SHH TP53 Mutant: wild 43:33 10:14 11:8 G3 MYC amplification: non-amplification 36:42 12:16 19:26 G4 Chromosome 11 loss: retain 18:17 3:6 22:17 Conventional MRI Features Cerebellar origin (yes: no) 3:42 71:5 9:69 3:42 2:18 19:5 5:23 3:6 2:19 13:6 6:39 3:36 Brainstem involvement (yes: no) 13:32 10:66 61:17 39:6 7:13 9:15 10:18 3:6 6:15 4:15 29:16 25:14 T1W images foci-hyperintensity (present: absent) 29:26 36:40 40:38 25:20 11:9 11:13 11:17 5:4 11:10 12:7 24:21 22:17 Enhancement degree on CE-T1W images (mild: obvious) 25:20 44:32 47:31 24:21 11:9 10:14 12:16 3:6 9:12 7:12 31:14 17:22 Enhancement appearance (homogeneous: heterogeneous) 17:28 36:40 33:45 18:27 12:8 6:18 13:15 2:7 7:14 8:11 17:28 13:26 T2W images foci-hypointensity (present: absent) 14:31 49:27 55:33 31:14 11:9 17:7 16:14 6:3 12:9 13:18 22:23 19:20 Hemorrhage (present: absent) 4:41 22:54 18:60 28:17 12:8 19:5 19:9 5:4 7:14 15:13 33:12 21:17 Cyst or necrosis (present: absent) 14:31 31:45 43:35 32:13 9:11 18:6 13:15 3:6 5:16 17:11 30:15 19:20 Dissemination/metastasis (yes: no) 12:33 18:58 23:55 19:28 5:15 7:17 8:20 2:7 4:17 5:23 12:33 5:34 Tumor volume (cm 3 ) 284.12 ± 13.22 198.11 ± 34.31 254.69 ± 35.51 336.80 ± 16.79 190.87 ± 34.21 223.01 ± 28.28 201.47 ± 21.55 240.38 ± 31.7 235.00 ± 12.70 191.10 ± 19.48 276.63 ± 23.14 304.26 ± 17.88 WNT = Wingless, SHH = Sonic Hedgehog, G3 = Group 3, and G4 = Group 4. Evaluation of different algorithmic models In this investigation, we conducted a comparative analysis of four distinct algorithms by training each on a dataset and then evaluating their respective performances on a validation dataset. The algorithms tested were VGG, Unet, Google Net, and ResNet-50. Based on the precision values with 95% confidence intervals (CIs), the following results were observed: VGG achieved an precision of 65.03% [95% CI: 47.28%-71.66%], Unet reached an precision of 72.00% [95% CI: 62.78%-82.33%], Google Net yielded an precision of 69.06% [95% CI: 55.43%-79.87%], and ResNet-50 led with an precision of 78.05% [95% CI: 69.00%-86.33%]. Given these outcomes, ResNet-50 emerged as the most proficient model in terms of classification accuracy and was subsequently chosen as the final model for our classification tasks. Table S1 and Figure S2 presented additional details. Results of initial stage: molecular subgrouping by MB-CNN The Dice score of MB-CNN was 0.91 ± 0.13. In the validation dataset, MB-CNN demonstrated notable accuracy in differentiating the four molecular subgroups of MB - WNT, SHH, G3, and G4. Comprehensive evaluation indicated that the model achieved an accuracy ranging from 74.68–77.84% and a precision rate between 73.85% and 78.90%. The model also showed a reliable range for Recall and F1 score, approximately spanning from 74.69–77.81% and 63.60–80.00%, respectively. In the independent external test dataset, the model's capability to distinguish between the MB molecular subgroups was further tested. Here, the model's accuracy was in the range of 76.29–78.71%. Table 2 , Fig. 3 A shows the classification effect of different subgroups in the validation and external test datasets. Table 2 Results of MB-CNN in validation and independent external test datasets and results of MB-CNN_TP53/MYC/Chr11 in different MB molecular subgroups. Subgroups Accuracy [95%CI] Recall [95%CI] Precision [95%CI] F1 Score [95%CI] Validation dataset WNT (n = 20) 74.68% [53.10% − 88.88%] 75.00% [55.60% − 93.30%] 78.90% [58.80% − 95.01%] 76.90% [58.80% − 90.09%] SHH (n = 24) 75.00% [55.11% − 88.25%] 74.69% [56.50% − 92.67%] 85.70% [58.88% − 95.00%] 80.00% [65.30% − 91.29%] Group 3 (n = 28) 76.29% [56.60% − 87.31%] 76.48% [57.10% − 90.00%] 75.00% [58.10% − 90.50%] 75.00% [60.00% − 86.20%] Group 4 (n = 9) 77.84% [45.33% − 93.70%] 77.81% [45.5% − 100.00%] 73.85% [55.00% − 84.67%] 63.60% [33.33% − 83.33%] Independent external test dataset WNT (n = 21) 76.29% [70.24% − 92.34%] 75.86% [60.29% − 91.44%] 78.75% [62.69% − 84.81%] 72.13% [36.60% − 72.84%] SHH (n = 19) 78.71% [73.14% − 94.28%] 82.76% [69.01% − 96.51%] 72.73% [57.53% − 87.92%] 77.42% [46.61% − 81.53%] Group 3 (n = 45) 77.90% [72.16% − 93.64%] 76.47% [62.21% − 90.73%] 78.79% [64.84% − 92.74%] 77.61% [42.84% − 75.86%] Group 4 (n = 39) 77.10% [71.20% − 92.99%] 65.63% [49.17% − 82.08%] 80.77% [65.62% − 95.92%] 72.41% [30.22% − 64.82%] AUC [95% CI] Accuracy% [95% CI] Sensitivity% [95% CI] Specificity% [95% CI] Validation dataset SHH (TP53 mutant status) (n = 24) 0.911 [0.820–0.960] 90.14% [87.63% − 96.42%] 78.62% [71.55% − 85.44%] 88.57% [78.02% − 94.43%] G3 (MYC amplification) (n = 28) 0.870 [0.760–0.940] 84.43% [77.19% − 93.53%] 84.32% [77.54% − 92.42%] 84.54% [77.61% − 90.43%] G4 (Chromosome 11 loss) (n = 9) 0.890 [0.800–0.980] 88.64% [73.42% − 99.44%] 80.54% [72.55% − 88.63%] 77.92% [69.41% − 90.85%] Independent external test dataset SHH (TP53 mutant status) ( n = 19 ) 0.930 [0.850–0.970] 91.27% [89.34% − 97.21%] 84.88% [80.23% − 86.54%] 89.66% [81.42% − 95.54%] G3 (MYC amplification) (n = 45) 0.850 [0.740–0.920] 83.56% [76.23% − 90.03%] 82.71% [75.44% − 89.98%] 85.34% [79.52% − 92.16%] G4 (Chromosome 11 loss) (n = 39) 0.880 [0.790–0.950] 86.79% [78.56% − 94.22%] 81.47% [77.58% − 89.36%] 82.38% [76.21% − 90.55%] SHH = Sonic Hedgehog, G3 = Group 3, and G4 = Group 4. AUC = Area under the curve. Results of the second stage: predicting prognostic related genetic signatures in MB with MB-CNN_TP53/MYC/Chr11 In our analysis, we subdivided MB into distinct subgroups within the development dataset, each characterized by prognostic related genetic signatures - TP53 gene mutation, MYC amplification, and chromosome 11 loss. Validation dataset analysis Identified by TP53 gene mutation, this subgroup showed superior classification capabilities. The SHH subgroup exhibited an AUC of 0.91 [95% CI: 0.82–0.96], a high accuracy of 90.14% [95% CI: 87.63%-96.42%]. Defined by MYC amplification, the G3 subgroup presented an AUC of 0.87 [95% CI: 0.76–0.94], with an accuracy of 84.43% [95% CI: 77.19%-93.53%]. Recognized by loss of chromosome 11, the G4 subgroup demonstrated an AUC of 0.89 [95% CI: 0.80–0.98], accuracy of 88.64% [95% CI: 73.42%-99.44%]. External test dataset analysis Exhibiting remarkable classification performance, the SHH subgroup in the external test dataset, indicated by the TP53 mutation, showed an AUC of 0.93 [95% CI: 0.85–0.97], an accuracy of 91.27% [95% CI: 89.34%-97.21%]. With MYC amplification, the G3 subgroup demonstrated an AUC of 0.85 [95% CI: 0.79–0.92], an accuracy of 83.56% [95% CI: 76.23%-90.03%]. Characterized by chromosome 11 loss, the G4 subgroup exhibited an AUC of 0.88 [95% CI: 0.79–0.95], an accuracy of 86.79% [95% CI: 78.56%-94.22%]. These results, highlighting the effectiveness of genetic signatures utilization for precise MB subgroup classification, are further detailed in Table 2 and Figs. 3 B, showcasing the second stage model's ability to discriminate among the subgroups based on their corresponding prognostic related genetic signatures. Figure 3 C and D show the AUC values and accuracy, sensitivity, and specificity of MB-CNN_TP53/MYC/Chr11 in validation and independent external test datasets. Results of additional analysis: logistic regression and hybrid model In our investigation, we developed a logistic regression model that integrated clinical parameters with features derived from MR radiographic assessment. Additionally, we synthesized a hybrid model by combining this logistic regression model with the outcomes from the MB-CNN. Both models underwent thorough evaluation on the independent external test dataset to ascertain their predictive efficacy. The logistic regression model, which utilized solely clinical and MR features, did not perform optimally. The model's accuracy was in the range of 58.28–65.22%, with precision rates varying from 53.85–66.67% (Fig. 4 A). In contrast, the hybrid model displayed enhancements in performance metrics. Conversely, the hybrid model demonstrated improvements across various performance indicators. Its accuracy ranged between 78.00% and 86.21%, precision varied from 75.86–88.89% (Fig. 4 B). The comparative analysis revealed that the accuracy of the MB-CNN model demonstrated a mean enhancement of 21.04% relative to the baseline logistic regression. Further, the hybrid model exhibited an average accuracy improvement of 5.58% over the MB-CNN. These improvements highlight the efficacy of combining MB-CNN with clinical information and conventional MR features. In a comparative analysis, the hybrid model was found to significantly outperform the logistic regression model ( P = .009) and was competitively aligned with MB-CNN ( P = .105), underscoring the advantageous potential of integrated approaches to enhance predictive precision within clinical settings. Table S2 details the classification effect of different models in the independent external test dataset. Figure 4 C presents the growth rates of the different models. Subgroup analyses We divided patients according to age into juvenile group (≤ 18 years) and adult group (> 18 years), and subgroup analysis was performed on data from the independent external test dataset. MB-CNN exhibited an accuracy of 68.97–85.67% and precision of 62.50–78.33% in the juvenile group. And MB-CNN achieved an accuracy of 66.67–81.82% and precision of 69.57–78.33% in the adult group. Recall, F1 scores and the increase rate of MB-CNN compared with logistic regression model and hybrid model compared with MB-CNN are presented in the Table S3 and Fig. 4 D. MB-CNN_TP53/MYC/Chr11 achieved AUC of 0.89 to 0.96, accuracy of 83.33–92.11% in juvenile group, and achieved AUC of 0.88 to 0.95, accuracy of 82.61–91.30% in adult group. The details of sensitivity and specificity are shown in Table S3 and Fig. 4 E. Discussion This study demonstrates the potential of deep learning models in classifying medulloblastoma subgroups and predicting prognostic genetic signatures, highlighting their utility in neuro-oncology diagnostics. Among the models tested, ResNet-50 outperformed others like VGG, Unet, and Google Net, achieving 78.71% accuracy and 80.77% precision on an external test dataset. Specialized models for predicting key genetic signatures-TP53 mutations in the SHH subgroup, MYC amplification in G3, and chromosome 11 loss in G4-showed high AUC values (up to 0.93 for TP53), supporting their clinical relevance in risk stratification. Furthermore, integrating DL outputs with clinical and MR data in a hybrid model improved diagnostic performance compared to traditional logistic regression, illustrating the power of combining advanced machine learning with conventional medical data to refine risk assessment and treatment planning. Medulloblastoma, the most common pediatric malignant brain tumor, is divided into four molecular subgroups, each with distinct prognostic implications( 12 , 20 , 21 ). Advances in gene expression and DNA methylation profiling have expanded our understanding of MB pathogenesis, revealing molecular signatures that can inform therapy( 4 , 16 , 17 , 20 ). However, identifying patients who may benefit from reduced intervention or those needing more intensive treatment remains a challenge( 22 ). Recognizing molecular subgroups is key to personalizing therapy, avoiding over- or under-treatment, and improving outcomes( 23 , 24 ). While the WNT subgroup has a favorable prognosis, its molecular signatures do not significantly alter the clinical approach, underscoring the complexity of MB diagnostics( 15 , 25 ). The integration of neuroimaging with machine learning holds significant promise for non-invasive molecular classification( 23 ), enabling more precise therapeutic strategies. This synergy not only aids in preoperative identification of MB subgroups but also enhances risk monitoring( 14 , 15 , 26 ). Traditional diagnostic approaches, relying heavily on neuroradiologists' experience, are limited by subjectivity and may misclassify subgroups( 27 ). Our study improves upon previous efforts by incorporating clinical, radiographic, and genetic data, significantly enhancing accuracy over prior models( 28 – 30 ). For example, while earlier work by Zhang et al. ( 15 )and Chen et al. ( 20 ) showed promise, they lacked comprehensive data integration and had limited robustness( 1 , 3 , 11 , 31 , 32 ).. By leveraging methylation and next-generation sequencing, our study introduces a more comprehensive and accurate approach to MB classification, offering a solid foundation for personalized treatment in clinical practice. Conclusion This study presents a deep learning framework utilizing MRI to predict four molecular subgroups and associated prognostic genetic signatures in medulloblastoma (MB). By integrating multi-semantic models and multidimensional data, we enhance the generalizability of AI in clinical contexts. While tissue specimens remain essential for diagnosis, this DL framework offers a cost-effective alternative or complement to traditional molecular risk assessments. Future work in imaging genomics and model deployment could expand personalized treatment strategies and inform clinical trial design. Abbreviations DL deep learning MB medulloblastoma WNT Wingless SHH Sonic Hedgehog G3 Group 3 G4 Group 4 AUC area under the receiver operating characteristic curve CIs confidence intervals Declarations Declarations Human Ethics and Consent to Participate declarations :: This multicenter retrospective study was approved by the institutional review board (Ethics committee approval No. 82172608). Competing interests: The authors declare that they have no competing interests. Fundings: This work was supported by the Natural Science Foundation of Beijing (L232079), National Science and Technology Major Project of the Ministry of Science and Technology of China (2022ZD0210100), National Natural Science Foundation of China (82273343, 82172608, 82101356, and 81902975), National Science Fund of Beijing for Distinguished Young Scholars (JQ24040), Beijing Nova Star Program (20220484058), Capital Medical University Fund for Excellent Young Scholars (KCB2304), and International Exchange and Cooperation Projects (2024-GJJL-10). Author Contribution LYN,LHL design of the work; LYN,LYW,LHL the acquisition, analysis, LJ interpretation of data; LYN the creation of new software used in the work; LYN,LYW have drafted the work; SHH,LYU,JT,QXG substantively revised it. Availability of data and materials: Codes are available via: https://github.com/YanongL/DL-model-code-for-MB-subgroups References Ramaswamy V, Remke M, Bouffet E, Bailey S, Clifford SC, Doz F, et al. Risk stratification of childhood medulloblastoma in the molecular era: the current consensus. Acta Neuropathol. 2016;131(6):821–31. Louis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol. 2021. Menyhart O, Gyorffy B. Molecular stratifications, biomarker candidates and new therapeutic options in current medulloblastoma treatment approaches. Cancer Metastasis Rev. 2020;39(1):211–33. Gajjar A, Robinson GW, Smith KS, Lin T, Merchant TE, Chintagumpala M, et al. Outcomes by Clinical and Molecular Features in Children With Medulloblastoma Treated With Risk-Adapted Therapy: Results of an International Phase III Trial (SJMB03). J Clin Oncol. 2021;39(7):822–35. Medulloblastoma. Nat Rev Dis Primers. 2019;5(1):12. Remke M, Ramaswamy V. WNT Medulloblastoma Limbo: How Low Can We Go? Clin Cancer Res. 2022;28(19):4161–3. Chen L, Li Y, Song Z, Xue S, Liu F, Chang X, et al. O-GlcNAcylation promotes cerebellum development and medulloblastoma oncogenesis via SHH signaling. Proc Natl Acad Sci U S A. 2022;119(34):e2202821119. Zhukova N, Ramaswamy V, Remke M, Pfaff E, Shih DJ, Martin DC, et al. Subgroup-specific prognostic implications of TP53 mutation in medulloblastoma. J Clin Oncol. 2013;31(23):2927–35. Shih DJ, Northcott PA, Remke M, Korshunov A, Ramaswamy V, Kool M, et al. Cytogenetic prognostication within medulloblastoma subgroups. J Clin Oncol. 2014;32(9):886–96. Fukuoka K, Kurihara J, Shofuda T, Kagawa N, Yamasaki K, Ando R, et al. Subtyping of Group 3/4 medulloblastoma as a potential prognostic biomarker among patients treated with reduced dose of craniospinal irradiation: a Japanese Pediatric Molecular Neuro-Oncology Group study. Acta Neuropathol Commun. 2023;11(1):153. Ramaswamy V, Remke M, Adamski J, Bartels U, Tabori U, Wang X, et al. Medulloblastoma subgroup-specific outcomes in irradiated children: who are the true high-risk patients? Neuro Oncol. 2016;18(2):291–7. Northcott PA, Buchhalter I, Morrissy AS, Hovestadt V, Weischenfeldt J, Ehrenberger T, et al. The whole-genome landscape of medulloblastoma subtypes. Nature. 2017;547(7663):311–7. Sharma T, Schwalbe EC, Williamson D, Sill M, Hovestadt V, Mynarek M, et al. Second-generation molecular subgrouping of medulloblastoma: an international meta-analysis of Group 3 and Group 4 subtypes. Acta Neuropathol. 2019;138(2):309–26. Chen X, Fan Z, Li KK, Wu G, Yang Z, Gao X, et al. Molecular subgrouping of medulloblastoma based on few-shot learning of multitasking using conventional MR images: a retrospective multicenter study. Neurooncol Adv. 2020;2(1):vdaa079. Zhang M, Wong SW, Wright JN, Wagner MW, Toescu S, Han M, et al. MRI Radiogenomics of Pediatric Medulloblastoma: A Multicenter Study. Radiology. 2022;304(2):406–16. Leary SES, Packer RJ, Li Y, Billups CA, Smith KS, Jaju A, et al. Efficacy of Carboplatin and Isotretinoin in Children With High-risk Medulloblastoma: A Randomized Clinical Trial From the Children's Oncology Group. JAMA Oncol. 2021;7(9):1313–21. Kumar R, Smith KS, Deng M, Terhune C, Robinson GW, Orr BA, et al. Clinical Outcomes and Patient-Matched Molecular Composition of Relapsed Medulloblastoma. J Clin Oncol. 2021;39(7):807–21. Savjani R. nnU-Net: Further Automating Biomedical Image Autosegmentation. Radiol Imaging Cancer. 2021;3(1):e209039. Konar D, Bhattacharyya S, Gandhi TK, Panigrahi BK, Jiang R. 3-D Quantum-Inspired Self-Supervised Tensor Network for Volumetric Segmentation of Medical Images. IEEE Trans Neural Netw Learn Syst. 2023;PP. Schwalbe EC, Lindsey JC, Nakjang S, Crosier S, Smith AJ, Hicks D, et al. Novel molecular subgroups for clinical classification and outcome prediction in childhood medulloblastoma: a cohort study. Lancet Oncol. 2017;18(7):958–71. Coltin H, Sundaresan L, Smith KS, Skowron P, Massimi L, Eberhart CG, et al. Subgroup and subtype-specific outcomes in adult medulloblastoma. Acta Neuropathol. 2021;142(5):859–71. Soomro TA, Zheng L, Afifi AJ, Ali A, Soomro S, Yin M, Gao J. Image Segmentation for MR Brain Tumor Detection Using Machine Learning: A Review. IEEE Rev Biomed Eng. 2023;16:70–90. Eran A, Ozturk A, Aygun N, Izbudak I. Medulloblastoma: atypical CT and MRI findings in children. Pediatr Radiol. 2010;40(7):1254–62. Dasgupta A, Gupta T, Maitre M, Kalra B, Chatterjee A, Krishnatry R, et al. Prognostic impact of semantic MRI features on survival outcomes in molecularly subtyped medulloblastoma. Strahlenther Onkol. 2022;198(3):291–303. Li J, Chen C, Fu R, Zhang Y, Fan Y, Xu J, Cen Y. Texture Analysis of T1-Weighted Contrast-Enhanced Magnetic Resonance Imaging Potentially Predicts Outcomes of Patients with Non-Wingless-Type/Non-Sonic Hedgehog Medulloblastoma. World Neurosurg. 2020;137:e27-e33. Yan J, Liu L, Wang W, Zhao Y, Li KK, Li K, et al. Radiomic Features From Multi-Parameter MRI Combined With Clinical Parameters Predict Molecular Subgroups in Patients With Medulloblastoma. Front Oncol. 2020;10:558162. Alharbi M, Mobark N, Bashawri Y, Abu Safieh L, Alowayn A, Aljelaify R, et al. Methylation Profiling of Medulloblastoma in a Clinical Setting Permits Sub-classification and Reveals New Outcome Predictions. Front Neurol. 2020;11:167. Magadza T, Viriri S. Deep Learning for Brain Tumor Segmentation: A Survey of State-of-the-Art. J Imaging. 2021;7(2). Richards BA, Lillicrap TP, Beaudoin P, Bengio Y, Bogacz R, Christensen A, et al. A deep learning framework for neuroscience. Nat Neurosci. 2019;22(11):1761–70. Jayachandran Preetha C, Meredig H, Brugnara G, Mahmutoglu MA, Foltyn M, Isensee F, et al. Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumour response assessment in neuro-oncology: a multicentre, retrospective cohort study. Lancet Digit Health. 2021;3(12):e784-e94. Waszak SM, Northcott PA, Buchhalter I, Robinson GW, Sutter C, Groebner S, et al. Spectrum and prevalence of genetic predisposition in medulloblastoma: a retrospective genetic study and prospective validation in a clinical trial cohort. Lancet Oncol. 2018;19(6):785–98. Luo Z, Xin D, Liao Y, Berry K, Ogurek S, Zhang F, et al. Loss of phosphatase CTDNEP1 potentiates aggressive medulloblastoma by triggering MYC amplification and genomic instability. Nat Commun. 2023;14(1):762. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterials12021.docx Cite Share Download PDF Status: Published Journal Publication published 15 Sep, 2025 Read the published version in Chinese Neurosurgical Journal → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6622165","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":463158660,"identity":"7486a397-2cb7-461e-b7d6-ac644c5f1bb4","order_by":0,"name":"Yanong Li","email":"","orcid":"","institution":"Beijing Tian Tan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yanong","middleName":"","lastName":"Li","suffix":""},{"id":463158661,"identity":"b96d31c3-8181-435f-aa8a-a8ce26ecddf7","order_by":1,"name":"Hailong Liu","email":"","orcid":"","institution":"Beijing Tian Tan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hailong","middleName":"","lastName":"Liu","suffix":""},{"id":463158662,"identity":"5e4d9301-e615-4b3f-8242-3774d6cfd63e","order_by":2,"name":"Yawei Liu","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yawei","middleName":"","lastName":"Liu","suffix":""},{"id":463158663,"identity":"ba275ad3-e259-4984-bbfe-de82d4d52bf3","order_by":3,"name":"Jane Li","email":"","orcid":"","institution":"Memorial Sloan Kettering Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Jane","middleName":"","lastName":"Li","suffix":""},{"id":463158664,"identity":"308423c0-e37b-4ed7-9de3-a2bf74cdb94e","order_by":4,"name":"Hiro Hiromichi Suzuki","email":"","orcid":"","institution":"National Cancer Center Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Hiro","middleName":"Hiromichi","lastName":"Suzuki","suffix":""},{"id":463158665,"identity":"f6e6f372-bbf8-4fe6-90e0-ba5aaa42fe88","order_by":5,"name":"Yaou Liu","email":"","orcid":"","institution":"Beijing Tian Tan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yaou","middleName":"","lastName":"Liu","suffix":""},{"id":463158669,"identity":"2b6269e7-114e-4c1c-bff6-d84f76ddd0d5","order_by":6,"name":"Tao Jiang","email":"","orcid":"","institution":"Beijing Tian Tan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Jiang","suffix":""},{"id":463158670,"identity":"49084ac2-34f6-4270-872c-7bf259a113e5","order_by":7,"name":"Xiaoguang Qiu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/0lEQVRIiWNgGAWjYDACZiBOgHE+NrAxMEiQooVxZgObBA9BLSjaeRsYCGvhO85jJvGgxsZuw/Gzh1/b7uCrs5duYHxc8Qu3FsnDPMYGCcfSkjecyUuzzj0DdJjMAWbDs324tRgc5jF8kMB2ONngQI6ZcW4byC8JbJKNPXi1GBxI+AfUcv6NmbElkVoMHyS2HbYzuJFj/JgRpqXhBz6/sBUbJPalJUjeeGPG2HuGTbLnRmKzYWMDbi185w9vk/zxzcae73yO8YefO47xs89IPviw4Q9uLQwHIFQi0Fg2YIwcA7IZGxgY2whrsQdi5g8MDDVQYXy2jIJRMApGwUgDAEdoU5LUSo9DAAAAAElFTkSuQmCC","orcid":"","institution":"Beijing Tian Tan Hospital","correspondingAuthor":true,"prefix":"","firstName":"Xiaoguang","middleName":"","lastName":"Qiu","suffix":""}],"badges":[],"createdAt":"2025-05-08 16:08:30","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6622165/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6622165/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s41016-025-00405-7","type":"published","date":"2025-09-15T15:57:33+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":83813178,"identity":"e7dad4cb-be0a-4c2e-9bda-61eadd7126fa","added_by":"auto","created_at":"2025-06-03 07:19:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":297701,"visible":true,"origin":"","legend":"\u003cp\u003eThe flowchart of overall study.\u003c/p\u003e","description":"","filename":"Figure10608.png","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/5b96115980056072f27726c4.png"},{"id":83812373,"identity":"ba7cc1a6-7d6a-44de-87b9-bde88c98b293","added_by":"auto","created_at":"2025-06-03 07:11:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":490413,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eA. \u003c/strong\u003e\u003c/em\u003eThe segmentation and initial stage classification architectures (MB-CNN) of DL model. \u003cem\u003e\u003cstrong\u003eB\u003c/strong\u003e\u003c/em\u003e. The performance of the second stage DL model (MB-CNN_TP53/MYC/Chr11) in stratifying patients with MB into different molecular subgroups (SHH, Group 3, and Group 4). \u003cem\u003e\u003cstrong\u003eC.\u003c/strong\u003e\u003c/em\u003e The additional analyses of this study (conventional clinical data were used to construct a logistic regression model and logistic regression model combined with MB-CNN output were used to construct a hybrid model). \u003cem\u003e\u003cstrong\u003eD.\u003c/strong\u003e\u003c/em\u003e The complementary sets of subgroup analyses (juvenile group vs. adult group).\u003c/p\u003e","description":"","filename":"FIGURE20423.png","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/e38c49e074aa64a86ce8bf7b.png"},{"id":83811324,"identity":"1b43daef-2a5c-4182-84b3-a4988116ab5f","added_by":"auto","created_at":"2025-06-03 07:03:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":253775,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eA. \u003c/strong\u003e\u003c/em\u003eInitial stage results of this study. The confusion matrix for the MB-CNN on validation and independent test datasets, used to assess the performance of a classification model. Each cell illustrates the relationship between actual and predicted categories, including Number (The large number in each cell indicates how many times each category was predicted as another category); Precision (Indicated by “P[rec] = X%\" in each cell, it measures the proportion of correct predictions out of all predictions for that category. Recall (Shown as \"Recall = X%\" in each cell, it measures the proportion of actual instances of a category that were correctly predicted).\u003cem\u003e\u003cstrong\u003e B.\u003c/strong\u003e\u003c/em\u003e Second stage results of this study. The receiver operator characteristic (ROC) curve of classification effects of second stage (MB-CNN_TP53/MYC/Chr11) on the validation datasetand independent test dataset. \u003cem\u003e\u003cstrong\u003eC\u003c/strong\u003e\u003c/em\u003e. Violin plot of area under the curve (AUC) values the MB-CNN_TP53/MYC/Chr11 to discriminate their corresponding gene signatures among different subgroups on validation and independent test dataset. \u003cem\u003e\u003cstrong\u003eD.\u003c/strong\u003e\u003c/em\u003e The dot plot of AUC, accuracy, sensitivity, and specificity of the second stage DL models (MB-CNN_TP53/MYC/Chr11) to discriminate their corresponding gene signatures among different subgroups on validation and independent test dataset.\u003c/p\u003e","description":"","filename":"Figure30608.png","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/27eba322b661f703e0cbca6e.png"},{"id":83811327,"identity":"3bbab39e-9b02-4603-a215-9f5939c127d6","added_by":"auto","created_at":"2025-06-03 07:03:12","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":177101,"visible":true,"origin":"","legend":"\u003cp\u003eAdditional analyses of this study. \u003cem\u003e\u003cstrong\u003eA\u003c/strong\u003e\u003c/em\u003e. The logistic regression model for confusion matrix results in the external test dataset. \u003cem\u003e\u003cstrong\u003eB\u003c/strong\u003e\u003c/em\u003e. The hybrid model for confusion matrix results in the external test dataset. \u003cem\u003e\u003cstrong\u003eC.\u003c/strong\u003e\u003c/em\u003e The performance metrics (Accuracy, F1 Score, Precision, and Recall) of the MB-CNN were systematically analyzed across various subgroups classifications of MB in independent test dataset. \u003cem\u003e\u003cstrong\u003eD. \u003c/strong\u003e\u003c/em\u003eAdditional analyses of this study. Confusion matrix results of the MB-CNN on the independent test dataset for the juvenile group (age ≤18 years) and adult group (age \u0026gt; 18 years). \u003cem\u003e\u003cstrong\u003eE.\u003c/strong\u003e\u003c/em\u003eThe receiver operator characteristic (ROC) curve of classification effects of second stage (MB-CNN_TP53/MYC/Chr11) on the independent test dataset for the juvenile group (age ≤ 18 years) and adult group (age \u0026gt;18 years) .\u003c/p\u003e","description":"","filename":"Figure40608.png","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/0c14261ae4c53fa0d1894b3d.png"},{"id":91889887,"identity":"a1279bbf-b3b7-421f-899e-4a946cddc768","added_by":"auto","created_at":"2025-09-22 16:03:15","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2425378,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/816d2df1-3fa7-40b2-b3b3-d1fb2b595a53.pdf"},{"id":83812375,"identity":"802573f4-bd0c-4618-8b87-41a037521ad9","added_by":"auto","created_at":"2025-06-03 07:11:12","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1288760,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials12021.docx","url":"https://assets-eu.researchsquare.com/files/rs-6622165/v1/6faa12857f5580a44e707da7.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Exploring Deep Learning and Hybrid Approaches in Molecular Subgrouping and Prognostic Related Genetic Signatures of Medulloblastoma","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMedulloblastoma (MB), the most prevalent malignant pediatric brain tumor in the posterior fossa(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e), is molecularly classified into four subgroups under the 2021 WHO framework: Wingless (WNT), Sonic Hedgehog (SHH), Group 3 (G3), and Group 4 (G4)1\u0026ndash;2(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Prognosis varies significantly across subgroups, with WNT tumors demonstrating\u0026thinsp;\u0026gt;\u0026thinsp;90% five-year survival, SHH subgroups exhibiting TP53 mutation-dependent heterogeneity, and MYC-amplified G3/G4 cases showing aggressive behavior and poor therapeutic response. This molecular stratification underpins current risk-adapted clinical management strategies(\u003cspan additionalcitationids=\"CR4 CR5 CR6 CR7 CR8 CR9 CR10\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDeep learning (DL) has revolutionized preoperative molecular prediction in gliomas, enabling accurate identification of biomarkers such as MGMT promoter methylation and IDH mutations. While preliminary DL applications in MB achieved 85% subgroup classification accuracy(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Critical gaps persist in current MB prognostic models, including insufficient cohort sizes for robust generalization, exclusion of high-risk genetic signatures (TP53 mutations, MYC amplification, chromosome 11 loss) in risk stratification frameworks, and underutilized opportunities to enhance predictive accuracy through multimodal clinical-MRI data integration(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo address these limitations, we develop a dual DL framework integrating preoperative MRI with methylation/NGS data. First, MB-CNN classifies tumors into molecular subgroups. Second, hybrid models (MB-CNN_TP53, MB-CNN_MYC, MB-CNN_Chr11) predict high-risk genetic signatures specific to each subgroup. Finally, a clinical-MRI fusion model combines radiomic features, conventional imaging biomarkers, and clinical variables to optimize prognostic accuracy. This approach bridges the translational gap between molecular subgroup classification and actionable genetic risk assessment.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and population\u003c/h2\u003e \u003cp\u003eThis study comprised two sequential stages. The initial stage developed a predictive DL model (MB-CNN) for molecular subgroups of MB (WNT, SHH, G3, and G4). The second stage constructed further DL models to predict prognostic related genetic signatures, TP53 mutation (MB-CNN_TP53), MYC amplification (MB-CNN_MYC), and chromosome 11 loss (MB-CNN_Chr11), within the SHH, G3, and G4 subgroups. Furthermore, additional analyses were performed. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the overall study flowchart. A total of 449 patients with MB in two independent medical institutes were consecutively enrolled (shown in \u003cb\u003eMethod S1\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis multicenter retrospective study was approved by the institutional review board (Ethics committee approval No. 82172608), with consent waived due to minimal risk. The inclusion criteria were as follows: (i) All patients diagnosed with MB confirmed between January 2015 and June 2023; (ii) necessary preoperative MR scans including axial T2-weighted (T2W), T1-weighted (T1W), and contrast-enhanced T1-weighted (CE-T1W) sequences. The exclusion criteria were as follows: (i) patients who lack the molecular subgroup and prognostic related genetic data; (ii) patients with prior neurosurgical procedures before MR scans; (iii) MR images with severe artifacts affecting radiological assessment.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eImaging preprocessing\u003c/h3\u003e\n\u003cp\u003eBrain MR images were performed with at 1.5- or 3.0-T magnet (Philips Healthcare, Siemens Healthineers, GE Healthcare, and Toshiba Medical Systems USA). The MRI scanning details are in \u003cb\u003eMethod S2\u003c/b\u003e. \u003cb\u003eFigure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e shows the proportion of manufactures across datasets. Preprocessing initiated with conversion to the NIfTI format (FSLv6.0, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://fsl.fmrib.ox.ac.uk/fsl/fslwiki/\u003c/span\u003e\u003cspan address=\"https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e for data consistency, followed by spatial resampling and intensity normalization within a range of -1 to 1 to reduce variability across scanners and improve comparability. \u003cb\u003eMethod S5\u003c/b\u003e contains comprehensive explanations of these methodological refinements to ensure data consistency and dependability across MR systems.\u003c/p\u003e\n\u003ch3\u003eBlinded tumor segmentation for ground truth generation\u003c/h3\u003e\n\u003cp\u003eWe, integrated methylation arrays, next-generation sequencing technologies (Whole Genome Sequencing (WGS) and Comparative Genomic Hybridization (CGH)), to accurately identify four molecular subgroups (presented in \u003cb\u003eMethod S4\u003c/b\u003e)(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFor medical image segmentation, the Unet architecture was selected for its proven effectiveness(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). Its Encoder-Decoder design extracts high-dimensional features via convolutional and max pooling layers, while up-convolution layers reconstruct segmentation maps for precise tumor identification.\u003c/p\u003e \u003cp\u003eThe segmentation model was developed in the development dataset of 325 patients\u0026rsquo; MR scans (T1W, CE-T1W, T2W), split into 80% training (n\u0026thinsp;=\u0026thinsp;260), 10% validation (n\u0026thinsp;=\u0026thinsp;33), and 10% testing (n\u0026thinsp;=\u0026thinsp;32). Training of the Unet model was conducted on the PyTorch (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pytorch.org/\u003c/span\u003e\u003cspan address=\"https://pytorch.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e over 100 epochs using the Adam optimizer with a cosine decay learning rate schedule. The Dice Loss function serves as the performance metric, with the model\u0026rsquo;s effectiveness periodically assessed via the Dice Score on the validation dataset.\u003c/p\u003e \u003cp\u003eTo establish a robust ground truth for the DL model, two neuroradiologists (Y.N.L. and D.C.., \u0026gt;\u0026thinsp;6 years experience) independently reviewed the segment results of Unet model. Low-quality images were resegmented using ITK-SNAP (version 3.8.0, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.itksnap.org\u003c/span\u003e\u003cspan address=\"http://www.itksnap.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e and refined by a senior neuroradiologist (Y.O.L., 20 years experience). Overlapping agreed regions served as the final ground truth for model development.\u003c/p\u003e\n\u003ch3\u003eConventional MR characteristics evaluation\u003c/h3\u003e\n\u003cp\u003eTwo experienced neuroradiologists (Y.N.L. and D.C.) independently evaluated conventional MR images characteristics across all datasets, including T1W foci-hyperintensity, T2W foci-hypointensity, and levels of enhancement on CE-T1W images (presented in \u003cb\u003eMethod S3\u003c/b\u003e). The agreement between readers was excellent (Kappa scores: 0.842\u0026ndash;0.864, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001). Discrepancies were adjudicated by a senior neuroradiologist (Y.O.L.).\u003c/p\u003e\n\u003ch3\u003eDL models training and validation\u003c/h3\u003e\n\u003cp\u003eIn this study, we developed MB-CNN and MB-CNN_TP53/MYC/Chr11 based on preoperative MR images (T1W, CE-T1W, and T2W). We modified the state-of-the-art 3D nnU-Net framework(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e) to segment the tumor and predict molecular subgroups and genetic signatures. The ResNet-50 architecture, characterized by its advanced deep residual learning strategy, incorporates a 50-layer deep structure designed to overcome the vanishing gradient issue common in extensive neural networks. Its residual blocks enhance gradient propagation throughout the network, enabling the efficient training of deeper neural models.\u003c/p\u003e \u003cp\u003eDuring the initial stage, for the MB-CNN project, ResNet-50 classified tumor molecular subgroups. Our study leveraged a dataset encompassing MR images from 325 patients, identical to the cohort used for segmentation model training. This dataset was methodically partitioned into training and validation datasets following 3:1 ratio (244:81), ensuring a randomized distribution. Additionally, a set of 124 independent external patients was designated as the test dataset. Each dataset entry encompasses a tumor segmentation map derived from the original MR scans, facilitated by the corresponding Unet model. The strategic dataset segmentation ensured equitable representation across the four molecular subgroups.\u003c/p\u003e \u003cp\u003eWhen the MB-CNN predicts the direct output result of WNT, and when the model predicts SHH, G3 or G4 subgroups, the prediction process will enter the second stage-three specialized two-class DL models (MB-CNN_TP53, MB-CNN_MYC, and MB-CNN_Chr11), based on ResNet-50 to assess prognostic genetic signatures. The models processed raw MR images and tumor segmentation maps produced by Unet. Prior to final classification through a fully connected layer, the images were resized to dimensions of 256x256x3 and underwent convolutional processing within the ResNet-50 framework. Subsequently, our model training utilized prognostic related genetic signatures status from SHH (n\u0026thinsp;=\u0026thinsp;54), G3 (n\u0026thinsp;=\u0026thinsp;78), and G4 (n\u0026thinsp;=\u0026thinsp;83) patient diagnoses. For validation, the model employed prognostic related genetic signatures identified within the validation dataset (SHH: n\u0026thinsp;=\u0026thinsp;14, G3: n\u0026thinsp;=\u0026thinsp;20, G4: n\u0026thinsp;=\u0026thinsp;21), with independent external test data serving for comprehensive testing (SHH: n\u0026thinsp;=\u0026thinsp;44, G3: n\u0026thinsp;=\u0026thinsp;23, G4: n\u0026thinsp;=\u0026thinsp;38).\u003c/p\u003e \u003cp\u003eThe classification model underwent 10 intensive training cycles on the PyTorch (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pytorch.org/\u003c/span\u003e\u003cspan address=\"https://pytorch.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e to enhance its accuracy in classification and prediction tasks. PyTorch was chosen for its support of dynamic computational graphs and GPU acceleration, essential for handling the ResNet-50 architecture and large datasets. To address class imbalance, the FocalLoss function prioritized hard-to-classify instances, enhancing sensitivity to underrepresented classes. The Adam optimizer adjusted learning rates for efficient convergence, refined by a cosine decay schedule.\u003c/p\u003e \u003cp\u003eTo evaluate the model\u0026rsquo;s generalizability and robustness, an independent dataset of 124 patients was tested, providing insights into the model's real-world applicability. For four-class classification, the MB-CNN model outputs a four-element vector of probabilities, each representing the likelihood of the input belonging to a specific category, ensured by the SoftMax activation function (\u003cb\u003eEq.\u0026nbsp;1\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:Softmax\\left({Z}_{i}\\right)=\\frac{{e}^{{Z}_{i}}}{{\\sum\\:}_{j}\\:{e}^{{Z}_{j}}}\\)\u003c/span\u003e \u003c/span\u003e (\u003cb\u003eEq.\u0026nbsp;1\u003c/b\u003e)\u003c/p\u003e \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{Z}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e is the value of the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e\u003csup\u003eth\u003c/sup\u003e element in vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{Z}\\)\u003c/span\u003e\u003c/span\u003e, and the denominator is the sum of applying the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}^{\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e function to all elements. Here, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{e}\\)\u003c/span\u003e\u003c/span\u003e is the base of the natural logarithm, approximately equal to 2.71828.\u003c/p\u003e \u003cp\u003eIn binary classification tasks, the model simplifies its output to a single probability value, indicating the likelihood of the input belonging to the \"positive\" class, a method that streamlines the output for binary tasks while maintaining predictive reliability.\u003c/p\u003e \u003cp\u003eThe DL model segmentation and classification architectures are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA. Further, we independently tested it in the independent external test dataset. The \u003cb\u003eMethod S6 and S7\u003c/b\u003e provide the training details of the segmentation and classification models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDL model evaluation\u003c/h2\u003e \u003cp\u003eThe predictive accuracy of the MB-CNN was rigorously assessed using a confusion matrix against the ground truth. The model produced probability scores for each subgroup, facilitating a probabilistic interpretation of results. For classification purposes, each instance was assigned to the molecular subgroup corresponding to the highest probability score. Accuracy and recall for each class were derived from the confusion matrix, offering insights into the model's performance on a per-class basis. These metrics were complemented by precision and the F1 score, which are particularly informative in scenarios of class imbalance, providing a balanced view of the model's predictive performance.\u003c/p\u003e \u003cp\u003eThe predictive accuracy of the second stage DL model (MB-CNN_TP53/MYC/Chr11) was assessed against ground truth via receiver operating characteristic (ROC) analysis. The model generated continuous probability scores for each subgroup. The optimal cutoff value for classifying subgroups was determined in the training dataset based on the Youden index, which maximizes the combined sensitivity and specificity. Subsequently, the molecular subgroup designation was assigned based on the highest predicted probability. The model's performance was evaluated using standard metrics, including the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Further, to ensure a robust evaluation, an independent external dataset comprising 105 patient samples (SHH: n\u0026thinsp;=\u0026thinsp;44, G3: n\u0026thinsp;=\u0026thinsp;23, G4: n\u0026thinsp;=\u0026thinsp;38) was utilized for testing. This external test step is critical for assessing the model's generalization capability and reliability in real-world scenarios, extending beyond the initial training and validation datasets.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAdditional analyses for clinical applicability\u003c/h3\u003e\n\u003cp\u003eTo further elucidate the potential application of the DL model in diverse clinical scenarios, we conducted two complementary sets of additional analyses. First, we constructed a logistic regression model using clinical information (age and gender) and conventional MR characteristics as a baseline for comparison. Subsequently, we built a hybrid model using logistic regression model incorporating the outputs of the MB-CNN. Ultimately, these analyses aim to evaluate the standalone performance of the MB-CNN. Additionally, they seek to assess the potential enhancive value provided by integrating ancillary inputs, such as clinical data and MR characteristics, into the MB-CNN framework. Finally, we divided the independent test dataset into adult group (\u0026gt;\u0026thinsp;18 years, n\u0026thinsp;=\u0026thinsp;26) and juvenile group (\u0026le;\u0026thinsp;18 years, n\u0026thinsp;=\u0026thinsp;98) according to age to compare MB-CNN and MB-CNN_TP53/MYC/Chr11 The models' performance was evaluated using confusion matrix (accuracy, recall, precision, and F1 score) and ROC curve (AUC, accuracy, sensitivity, and specificity).\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eWe used the Statistical Package for the Social Sciences (SPSS) software (version 22, IBM, USA) and Python (version 3.6, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://www.python.org\u003c/span\u003e\u003cspan address=\"http://www.python.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e for statistical analyses. Categorical variables were displayed as frequency and percentages and tested using Pearson's Chi-squared test or Fisher's exact test. Continuous variables were displayed as mean and standard deviation (SD) and were tested using a two-sample t-test or Mann\u0026ndash;Whitney U test (the choice between the t-test and Mann\u0026ndash;Whitney U test was based on normality tests). A two-sided \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.05 was considered statistically significant. \u003cb\u003eMethod S8\u003c/b\u003e shows more details of the Statistical analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eCode Availability\u003c/h2\u003e \u003cp\u003eCodes are available via: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/YanongL/DL-model-code-for-MB-subgroups\u003c/span\u003e\u003cspan address=\"https://github.com/YanongL/DL-model-code-for-MB-subgroups\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePatient characteristics\u003c/h2\u003e \u003cp\u003eOur study cohort consecutive included 449 patients diagnosed with MB, comprising 249 males and 200 females [mean age\u0026thinsp;=\u0026thinsp;13.63\u0026thinsp;\u0026plusmn;\u0026thinsp;2.47]. The age distribution revealed three distinct groups: 14 patients younger than 3 years [mean age\u0026thinsp;=\u0026thinsp;1.86\u0026thinsp;\u0026plusmn;\u0026thinsp;1.06], 243 patients aged 3\u0026ndash;18 years [mean age\u0026thinsp;=\u0026thinsp;11.30\u0026thinsp;\u0026plusmn;\u0026thinsp;3.12], and 68 adults aged 18\u0026ndash;49 years [mean age\u0026thinsp;=\u0026thinsp;28.67\u0026thinsp;\u0026plusmn;\u0026thinsp;4.33]. Detailed demographic information, genetic profiling, and MR characteristics of the study population are provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe main demographic and conventional MR features of the training dataset, validation dataset, and the independent external test dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"21\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c13\" colnum=\"13\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c14\" colnum=\"14\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c15\" colnum=\"15\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c16\" colnum=\"16\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c17\" colnum=\"17\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c18\" colnum=\"18\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c19\" colnum=\"19\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c20\" colnum=\"20\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c21\" colnum=\"21\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"13\" nameend=\"c14\" namest=\"c2\"\u003e \u003cp\u003eDevelopment Set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003eIndependent External Test Dataset\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003eTraining dataset (n\u0026thinsp;=\u0026thinsp;244)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c14\" namest=\"c7\"\u003e \u003cp\u003eValidation dataset (n\u0026thinsp;=\u0026thinsp;81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;124)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMolecular subgroups\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWNT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSHH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003eG3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eG4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eWNT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eSHH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c13\" namest=\"c11\"\u003e \u003cp\u003eG3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c14\"\u003e \u003cp\u003eG4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c16\" namest=\"c15\"\u003e \u003cp\u003eWNT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c18\" namest=\"c17\"\u003e \u003cp\u003eSHH\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c20\" namest=\"c19\"\u003e \u003cp\u003eG3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c21\"\u003e \u003cp\u003eG4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCount, n\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003e78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c13\" namest=\"c11\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c14\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c16\" namest=\"c15\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c18\" namest=\"c17\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c20\" namest=\"c19\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c21\"\u003e \u003cp\u003e39\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"21\" nameend=\"c21\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDemographic\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD), y\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003e14.42\u0026thinsp;\u0026plusmn;\u0026thinsp;1.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c14\" namest=\"c7\"\u003e \u003cp\u003e13.11\u0026thinsp;\u0026plusmn;\u0026thinsp;2.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003e12.53\u0026thinsp;\u0026plusmn;\u0026thinsp;3.21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGender, n\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"20\" nameend=\"c21\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale: Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22:23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e40:36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e29:49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e12:33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e \u003cp\u003e8:12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e \u003cp\u003e15:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e8:20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e2:7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e9:12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e10:19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e28:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e17:12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSHH\u003c/b\u003e\u003c/p\u003e \u003cp\u003eTP53 Mutant: wild\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003e43:33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c14\" namest=\"c7\"\u003e \u003cp\u003e10:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003e11:8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG3\u003c/b\u003e\u003c/p\u003e \u003cp\u003eMYC amplification: non-amplification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003e36:42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c14\" namest=\"c7\"\u003e \u003cp\u003e12:16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003e19:26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eG4\u003c/b\u003e\u003c/p\u003e \u003cp\u003eChromosome 11 loss: retain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e \u003cp\u003e18:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c14\" namest=\"c7\"\u003e \u003cp\u003e3:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c21\" namest=\"c15\"\u003e \u003cp\u003e22:17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"21\" nameend=\"c21\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eConventional MRI Features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCerebellar origin (yes: no)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3:42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71:5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9:69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e3:42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e2:18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e19:5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e5:23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e3:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e2:19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e13:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e6:39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e3:36\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBrainstem involvement (yes: no)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13:32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10:66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e61:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e39:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e7:13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e9:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e10:18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e3:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e6:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e4:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e29:16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e25:14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT1W images foci-hyperintensity (present: absent)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29:26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36:40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e40:38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e25:20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e11:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e11:13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e11:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e5:4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e11:10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e12:7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e24:21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e22:17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhancement degree on CE-T1W images\u003c/p\u003e \u003cp\u003e(mild: obvious)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25:20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44:32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e47:31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e24:21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e11:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e10:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e12:16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e3:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e9:12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e7:12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e31:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e17:22\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnhancement appearance (homogeneous: heterogeneous)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17:28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36:40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e33:45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e18:27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e12:8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e6:18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e13:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e2:7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e7:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e8:11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e17:28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e13:26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eT2W images foci-hypointensity (present: absent)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14:31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49:27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e55:33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e31:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e11:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e17:7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e16:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e6:3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e12:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e13:18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e22:23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e19:20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHemorrhage (present: absent)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4:41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22:54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18:60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e28:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e12:8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e19:5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e19:9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e5:4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e7:14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e15:13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e33:12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e21:17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCyst or necrosis (present: absent)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14:31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31:45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e43:35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e32:13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e9:11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e18:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e13:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e3:6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e5:16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e17:11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e30:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e19:20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDissemination/metastasis (yes: no)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12:33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18:58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23:55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e19:28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e5:15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e7:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e8:20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e2:7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e4:17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e5:23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e12:33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e5:34\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTumor volume (cm\u003csup\u003e3\u003c/sup\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e284.12\u0026thinsp;\u0026plusmn;\u0026thinsp;13.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e198.11\u0026thinsp;\u0026plusmn;\u0026thinsp;34.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e254.69\u0026thinsp;\u0026plusmn;\u0026thinsp;35.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003e336.80\u0026thinsp;\u0026plusmn;\u0026thinsp;16.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c8\" namest=\"c7\"\u003e \u003cp\u003e190.87\u0026thinsp;\u0026plusmn;\u0026thinsp;34.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c11\" namest=\"c9\"\u003e \u003cp\u003e223.01\u0026thinsp;\u0026plusmn;\u0026thinsp;28.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c12\"\u003e \u003cp\u003e201.47\u0026thinsp;\u0026plusmn;\u0026thinsp;21.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c14\" namest=\"c13\"\u003e \u003cp\u003e240.38\u0026thinsp;\u0026plusmn;\u0026thinsp;31.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c15\"\u003e \u003cp\u003e235.00\u0026thinsp;\u0026plusmn;\u0026thinsp;12.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c17\" namest=\"c16\"\u003e \u003cp\u003e191.10\u0026thinsp;\u0026plusmn;\u0026thinsp;19.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c19\" namest=\"c18\"\u003e \u003cp\u003e276.63\u0026thinsp;\u0026plusmn;\u0026thinsp;23.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c21\" namest=\"c20\"\u003e \u003cp\u003e304.26\u0026thinsp;\u0026plusmn;\u0026thinsp;17.88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"21\" nameend=\"c21\" namest=\"c1\"\u003e \u003cp\u003eWNT\u0026thinsp;=\u0026thinsp;Wingless, SHH\u0026thinsp;=\u0026thinsp;Sonic Hedgehog, G3\u0026thinsp;=\u0026thinsp;Group 3, and G4\u0026thinsp;=\u0026thinsp;Group 4.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation of different algorithmic models\u003c/h2\u003e \u003cp\u003eIn this investigation, we conducted a comparative analysis of four distinct algorithms by training each on a dataset and then evaluating their respective performances on a validation dataset. The algorithms tested were VGG, Unet, Google Net, and ResNet-50. Based on the precision values with 95% confidence intervals (CIs), the following results were observed: VGG achieved an precision of 65.03% [95% CI: 47.28%-71.66%], Unet reached an precision of 72.00% [95% CI: 62.78%-82.33%], Google Net yielded an precision of 69.06% [95% CI: 55.43%-79.87%], and ResNet-50 led with an precision of 78.05% [95% CI: 69.00%-86.33%]. Given these outcomes, ResNet-50 emerged as the most proficient model in terms of classification accuracy and was subsequently chosen as the final model for our classification tasks. \u003cb\u003eTable \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e and \u003cb\u003eFigure S2\u003c/b\u003e presented additional details.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eResults of initial stage: molecular subgrouping by MB-CNN\u003c/h2\u003e \u003cp\u003eThe Dice score of MB-CNN was 0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.13. In the validation dataset, MB-CNN demonstrated notable accuracy in differentiating the four molecular subgroups of MB - WNT, SHH, G3, and G4. Comprehensive evaluation indicated that the model achieved an accuracy ranging from 74.68\u0026ndash;77.84% and a precision rate between 73.85% and 78.90%. The model also showed a reliable range for Recall and F1 score, approximately spanning from 74.69\u0026ndash;77.81% and 63.60\u0026ndash;80.00%, respectively.\u003c/p\u003e \u003cp\u003eIn the independent external test dataset, the model's capability to distinguish between the MB molecular subgroups was further tested. Here, the model's accuracy was in the range of 76.29\u0026ndash;78.71%. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA shows the classification effect of different subgroups in the validation and external test datasets.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults of MB-CNN in validation and independent external test datasets and results of MB-CNN_TP53/MYC/Chr11 in different MB molecular subgroups.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSubgroups\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy [95%CI]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall [95%CI]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision [95%CI]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1 Score [95%CI]\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003eValidation dataset\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWNT (n\u0026thinsp;=\u0026thinsp;20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e74.68% [53.10% \u0026minus;\u0026thinsp;88.88%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e75.00% [55.60% \u0026minus;\u0026thinsp;93.30%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.90% [58.80% \u0026minus;\u0026thinsp;95.01%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e76.90% [58.80% \u0026minus;\u0026thinsp;90.09%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSHH (n\u0026thinsp;=\u0026thinsp;24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75.00% [55.11% \u0026minus;\u0026thinsp;88.25%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e74.69% [56.50% \u0026minus;\u0026thinsp;92.67%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e85.70% [58.88% \u0026minus;\u0026thinsp;95.00%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e80.00% [65.30% \u0026minus;\u0026thinsp;91.29%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 3 (n\u0026thinsp;=\u0026thinsp;28)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e76.29% [56.60% \u0026minus;\u0026thinsp;87.31%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76.48% [57.10% \u0026minus;\u0026thinsp;90.00%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e75.00% [58.10% \u0026minus;\u0026thinsp;90.50%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e75.00% [60.00% \u0026minus;\u0026thinsp;86.20%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 4 (n\u0026thinsp;=\u0026thinsp;9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77.84% [45.33% \u0026minus;\u0026thinsp;93.70%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e77.81% [45.5% \u0026minus;\u0026thinsp;100.00%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e73.85% [55.00% \u0026minus;\u0026thinsp;84.67%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63.60% [33.33% \u0026minus;\u0026thinsp;83.33%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIndependent external test dataset\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWNT (n\u0026thinsp;=\u0026thinsp;21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e76.29% [70.24% \u0026minus;\u0026thinsp;92.34%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e75.86% [60.29% \u0026minus;\u0026thinsp;91.44%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.75% [62.69% \u0026minus;\u0026thinsp;84.81%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e72.13% [36.60% \u0026minus;\u0026thinsp;72.84%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSHH (n\u0026thinsp;=\u0026thinsp;19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e78.71% [73.14% \u0026minus;\u0026thinsp;94.28%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e82.76% [69.01% \u0026minus;\u0026thinsp;96.51%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e72.73% [57.53% \u0026minus;\u0026thinsp;87.92%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.42% [46.61% \u0026minus;\u0026thinsp;81.53%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 3 (n\u0026thinsp;=\u0026thinsp;45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77.90% [72.16% \u0026minus;\u0026thinsp;93.64%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76.47% [62.21% \u0026minus;\u0026thinsp;90.73%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.79% [64.84% \u0026minus;\u0026thinsp;92.74%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.61% [42.84% \u0026minus;\u0026thinsp;75.86%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 4 (n\u0026thinsp;=\u0026thinsp;39)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77.10% [71.20% \u0026minus;\u0026thinsp;92.99%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e65.63% [49.17% \u0026minus;\u0026thinsp;82.08%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e80.77% [65.62% \u0026minus;\u0026thinsp;95.92%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e72.41% [30.22% \u0026minus;\u0026thinsp;64.82%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAUC [95% CI]\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003eAccuracy% [95% CI]\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eSensitivity% [95% CI]\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eSpecificity% [95% CI]\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eValidation dataset\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSHH (TP53 mutant status) (n\u0026thinsp;=\u0026thinsp;24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.911 [0.820\u0026ndash;0.960]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e90.14% [87.63% \u0026minus;\u0026thinsp;96.42%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.62% [71.55% \u0026minus;\u0026thinsp;85.44%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88.57% [78.02% \u0026minus;\u0026thinsp;94.43%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eG3 (MYC amplification) (n\u0026thinsp;=\u0026thinsp;28)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.870 [0.760\u0026ndash;0.940]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84.43% [77.19% \u0026minus;\u0026thinsp;93.53%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e84.32% [77.54% \u0026minus;\u0026thinsp;92.42%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e84.54% [77.61% \u0026minus;\u0026thinsp;90.43%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eG4 (Chromosome 11 loss) (n\u0026thinsp;=\u0026thinsp;9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.890 [0.800\u0026ndash;0.980]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e88.64% [73.42% \u0026minus;\u0026thinsp;99.44%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e80.54% [72.55% \u0026minus;\u0026thinsp;88.63%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e77.92% [69.41% \u0026minus;\u0026thinsp;90.85%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIndependent external test dataset\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSHH (TP53 mutant status) \u003cb\u003e(\u003c/b\u003en\u0026thinsp;=\u0026thinsp;19\u003cb\u003e)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.930 [0.850\u0026ndash;0.970]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91.27% [89.34% \u0026minus;\u0026thinsp;97.21%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e84.88% [80.23% \u0026minus;\u0026thinsp;86.54%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e89.66% [81.42% \u0026minus;\u0026thinsp;95.54%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eG3 (MYC amplification) (n\u0026thinsp;=\u0026thinsp;45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.850 [0.740\u0026ndash;0.920]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e83.56% [76.23% \u0026minus;\u0026thinsp;90.03%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e82.71% [75.44% \u0026minus;\u0026thinsp;89.98%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e85.34% [79.52% \u0026minus;\u0026thinsp;92.16%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eG4 (Chromosome 11 loss) (n\u0026thinsp;=\u0026thinsp;39)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.880 [0.790\u0026ndash;0.950]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e86.79% [78.56% \u0026minus;\u0026thinsp;94.22%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e81.47% [77.58% \u0026minus;\u0026thinsp;89.36%]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e82.38% [76.21% \u0026minus;\u0026thinsp;90.55%]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003eSHH\u0026thinsp;=\u0026thinsp;Sonic Hedgehog, G3\u0026thinsp;=\u0026thinsp;Group 3, and G4\u0026thinsp;=\u0026thinsp;Group 4.\u003c/p\u003e \u003cp\u003eAUC\u0026thinsp;=\u0026thinsp;Area under the curve.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eResults of the second stage: predicting prognostic related genetic signatures in MB with MB-CNN_TP53/MYC/Chr11\u003c/h2\u003e \u003cp\u003eIn our analysis, we subdivided MB into distinct subgroups within the development dataset, each characterized by prognostic related genetic signatures - TP53 gene mutation, MYC amplification, and chromosome 11 loss.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eValidation dataset analysis\u003c/h2\u003e \u003cp\u003eIdentified by TP53 gene mutation, this subgroup showed superior classification capabilities. The SHH subgroup exhibited an AUC of 0.91 [95% CI: 0.82\u0026ndash;0.96], a high accuracy of 90.14% [95% CI: 87.63%-96.42%].\u003c/p\u003e \u003cp\u003eDefined by MYC amplification, the G3 subgroup presented an AUC of 0.87 [95% CI: 0.76\u0026ndash;0.94], with an accuracy of 84.43% [95% CI: 77.19%-93.53%].\u003c/p\u003e \u003cp\u003eRecognized by loss of chromosome 11, the G4 subgroup demonstrated an AUC of 0.89 [95% CI: 0.80\u0026ndash;0.98], accuracy of 88.64% [95% CI: 73.42%-99.44%].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eExternal test dataset analysis\u003c/h2\u003e \u003cp\u003eExhibiting remarkable classification performance, the SHH subgroup in the external test dataset, indicated by the TP53 mutation, showed an AUC of 0.93 [95% CI: 0.85\u0026ndash;0.97], an accuracy of 91.27% [95% CI: 89.34%-97.21%].\u003c/p\u003e \u003cp\u003eWith MYC amplification, the G3 subgroup demonstrated an AUC of 0.85 [95% CI: 0.79\u0026ndash;0.92], an accuracy of 83.56% [95% CI: 76.23%-90.03%].\u003c/p\u003e \u003cp\u003eCharacterized by chromosome 11 loss, the G4 subgroup exhibited an AUC of 0.88 [95% CI: 0.79\u0026ndash;0.95], an accuracy of 86.79% [95% CI: 78.56%-94.22%]. These results, highlighting the effectiveness of genetic signatures utilization for precise MB subgroup classification, are further detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB, showcasing the second stage model's ability to discriminate among the subgroups based on their corresponding prognostic related genetic signatures. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC and \u003cb\u003eD\u003c/b\u003e show the AUC values and accuracy, sensitivity, and specificity of MB-CNN_TP53/MYC/Chr11 in validation and independent external test datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eResults of additional analysis: logistic regression and hybrid model\u003c/h2\u003e \u003cp\u003eIn our investigation, we developed a logistic regression model that integrated clinical parameters with features derived from MR radiographic assessment. Additionally, we synthesized a hybrid model by combining this logistic regression model with the outcomes from the MB-CNN. Both models underwent thorough evaluation on the independent external test dataset to ascertain their predictive efficacy.\u003c/p\u003e \u003cp\u003eThe logistic regression model, which utilized solely clinical and MR features, did not perform optimally. The model's accuracy was in the range of 58.28\u0026ndash;65.22%, with precision rates varying from 53.85\u0026ndash;66.67% (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn contrast, the hybrid model displayed enhancements in performance metrics. Conversely, the hybrid model demonstrated improvements across various performance indicators. Its accuracy ranged between 78.00% and 86.21%, precision varied from 75.86\u0026ndash;88.89% (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). The comparative analysis revealed that the accuracy of the MB-CNN model demonstrated a mean enhancement of 21.04% relative to the baseline logistic regression. Further, the hybrid model exhibited an average accuracy improvement of 5.58% over the MB-CNN. These improvements highlight the efficacy of combining MB-CNN with clinical information and conventional MR features.\u003c/p\u003e \u003cp\u003eIn a comparative analysis, the hybrid model was found to significantly outperform the logistic regression model (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.009) and was competitively aligned with MB-CNN (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.105), underscoring the advantageous potential of integrated approaches to enhance predictive precision within clinical settings. \u003cb\u003eTable S2\u003c/b\u003e details the classification effect of different models in the independent external test dataset. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC presents the growth rates of the different models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eSubgroup analyses\u003c/h2\u003e \u003cp\u003eWe divided patients according to age into juvenile group (\u0026le;\u0026thinsp;18 years) and adult group (\u0026gt;\u0026thinsp;18 years), and subgroup analysis was performed on data from the independent external test dataset. MB-CNN exhibited an accuracy of 68.97\u0026ndash;85.67% and precision of 62.50\u0026ndash;78.33% in the juvenile group. And MB-CNN achieved an accuracy of 66.67\u0026ndash;81.82% and precision of 69.57\u0026ndash;78.33% in the adult group. Recall, F1 scores and the increase rate of MB-CNN compared with logistic regression model and hybrid model compared with MB-CNN are presented in the \u003cb\u003eTable S3\u003c/b\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eD. MB-CNN_TP53/MYC/Chr11 achieved AUC of 0.89 to 0.96, accuracy of 83.33\u0026ndash;92.11% in juvenile group, and achieved AUC of 0.88 to 0.95, accuracy of 82.61\u0026ndash;91.30% in adult group. The details of sensitivity and specificity are shown in \u003cb\u003eTable S3\u003c/b\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eE.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrates the potential of deep learning models in classifying medulloblastoma subgroups and predicting prognostic genetic signatures, highlighting their utility in neuro-oncology diagnostics. Among the models tested, ResNet-50 outperformed others like VGG, Unet, and Google Net, achieving 78.71% accuracy and 80.77% precision on an external test dataset. Specialized models for predicting key genetic signatures-TP53 mutations in the SHH subgroup, MYC amplification in G3, and chromosome 11 loss in G4-showed high AUC values (up to 0.93 for TP53), supporting their clinical relevance in risk stratification. Furthermore, integrating DL outputs with clinical and MR data in a hybrid model improved diagnostic performance compared to traditional logistic regression, illustrating the power of combining advanced machine learning with conventional medical data to refine risk assessment and treatment planning.\u003c/p\u003e \u003cp\u003eMedulloblastoma, the most common pediatric malignant brain tumor, is divided into four molecular subgroups, each with distinct prognostic implications(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Advances in gene expression and DNA methylation profiling have expanded our understanding of MB pathogenesis, revealing molecular signatures that can inform therapy(\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e). However, identifying patients who may benefit from reduced intervention or those needing more intensive treatment remains a challenge(\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). Recognizing molecular subgroups is key to personalizing therapy, avoiding over- or under-treatment, and improving outcomes(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). While the WNT subgroup has a favorable prognosis, its molecular signatures do not significantly alter the clinical approach, underscoring the complexity of MB diagnostics(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe integration of neuroimaging with machine learning holds significant promise for non-invasive molecular classification(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e), enabling more precise therapeutic strategies. This synergy not only aids in preoperative identification of MB subgroups but also enhances risk monitoring(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). Traditional diagnostic approaches, relying heavily on neuroradiologists' experience, are limited by subjectivity and may misclassify subgroups(\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e). Our study improves upon previous efforts by incorporating clinical, radiographic, and genetic data, significantly enhancing accuracy over prior models(\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e). For example, while earlier work by Zhang et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e)and Chen et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e) showed promise, they lacked comprehensive data integration and had limited robustness(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e).. By leveraging methylation and next-generation sequencing, our study introduces a more comprehensive and accurate approach to MB classification, offering a solid foundation for personalized treatment in clinical practice.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study presents a deep learning framework utilizing MRI to predict four molecular subgroups and associated prognostic genetic signatures in medulloblastoma (MB). By integrating multi-semantic models and multidimensional data, we enhance the generalizability of AI in clinical contexts. While tissue specimens remain essential for diagnosis, this DL framework offers a cost-effective alternative or complement to traditional molecular risk assessments. Future work in imaging genomics and model deployment could expand personalized treatment strategies and inform clinical trial design.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003edeep learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emedulloblastoma\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eWNT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWingless\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSHH\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSonic Hedgehog\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eG3\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGroup 3\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eG4\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGroup 4\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003earea under the receiver operating characteristic curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCIs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003econfidence intervals\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eDeclarations\u003c/h2\u003e \u003cp\u003e \u003cb\u003eHuman Ethics and Consent to Participate declarations\u003c/b\u003e::\u003c/p\u003e \u003cp\u003eThis multicenter retrospective study was approved by the institutional review board (Ethics committee approval No. 82172608).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eCompeting interests:\u003c/strong\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFundings:\u003c/h2\u003e \u003cp\u003eThis work was supported by the Natural Science Foundation of Beijing (L232079), National Science and Technology Major Project of the Ministry of Science and Technology of China (2022ZD0210100), National Natural Science Foundation of China (82273343, 82172608, 82101356, and 81902975), National Science Fund of Beijing for Distinguished Young Scholars (JQ24040), Beijing Nova Star Program (20220484058), Capital Medical University Fund for Excellent Young Scholars (KCB2304), and International Exchange and Cooperation Projects (2024-GJJL-10).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eLYN,LHL design of the work; LYN,LYW,LHL the acquisition, analysis, LJ interpretation of data; LYN the creation of new software used in the work; LYN,LYW have drafted the work; SHH,LYU,JT,QXG substantively revised it.\u003c/p\u003e\u003ch2\u003eAvailability of data and materials:\u003c/h2\u003e \u003cp\u003eCodes are available via: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/YanongL/DL-model-code-for-MB-subgroups\u003c/span\u003e\u003cspan address=\"https://github.com/YanongL/DL-model-code-for-MB-subgroups\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eRamaswamy V, Remke M, Bouffet E, Bailey S, Clifford SC, Doz F, et al. Risk stratification of childhood medulloblastoma in the molecular era: the current consensus. Acta Neuropathol. 2016;131(6):821\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLouis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol. 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenyhart O, Gyorffy B. Molecular stratifications, biomarker candidates and new therapeutic options in current medulloblastoma treatment approaches. Cancer Metastasis Rev. 2020;39(1):211\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGajjar A, Robinson GW, Smith KS, Lin T, Merchant TE, Chintagumpala M, et al. Outcomes by Clinical and Molecular Features in Children With Medulloblastoma Treated With Risk-Adapted Therapy: Results of an International Phase III Trial (SJMB03). J Clin Oncol. 2021;39(7):822\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMedulloblastoma. Nat Rev Dis Primers. 2019;5(1):12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRemke M, Ramaswamy V. WNT Medulloblastoma Limbo: How Low Can We Go? Clin Cancer Res. 2022;28(19):4161\u0026ndash;3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen L, Li Y, Song Z, Xue S, Liu F, Chang X, et al. O-GlcNAcylation promotes cerebellum development and medulloblastoma oncogenesis via SHH signaling. Proc Natl Acad Sci U S A. 2022;119(34):e2202821119.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhukova N, Ramaswamy V, Remke M, Pfaff E, Shih DJ, Martin DC, et al. Subgroup-specific prognostic implications of TP53 mutation in medulloblastoma. J Clin Oncol. 2013;31(23):2927\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShih DJ, Northcott PA, Remke M, Korshunov A, Ramaswamy V, Kool M, et al. Cytogenetic prognostication within medulloblastoma subgroups. J Clin Oncol. 2014;32(9):886\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFukuoka K, Kurihara J, Shofuda T, Kagawa N, Yamasaki K, Ando R, et al. Subtyping of Group 3/4 medulloblastoma as a potential prognostic biomarker among patients treated with reduced dose of craniospinal irradiation: a Japanese Pediatric Molecular Neuro-Oncology Group study. Acta Neuropathol Commun. 2023;11(1):153.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamaswamy V, Remke M, Adamski J, Bartels U, Tabori U, Wang X, et al. Medulloblastoma subgroup-specific outcomes in irradiated children: who are the true high-risk patients? Neuro Oncol. 2016;18(2):291\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNorthcott PA, Buchhalter I, Morrissy AS, Hovestadt V, Weischenfeldt J, Ehrenberger T, et al. The whole-genome landscape of medulloblastoma subtypes. Nature. 2017;547(7663):311\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharma T, Schwalbe EC, Williamson D, Sill M, Hovestadt V, Mynarek M, et al. Second-generation molecular subgrouping of medulloblastoma: an international meta-analysis of Group 3 and Group 4 subtypes. Acta Neuropathol. 2019;138(2):309\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X, Fan Z, Li KK, Wu G, Yang Z, Gao X, et al. Molecular subgrouping of medulloblastoma based on few-shot learning of multitasking using conventional MR images: a retrospective multicenter study. Neurooncol Adv. 2020;2(1):vdaa079.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang M, Wong SW, Wright JN, Wagner MW, Toescu S, Han M, et al. MRI Radiogenomics of Pediatric Medulloblastoma: A Multicenter Study. Radiology. 2022;304(2):406\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeary SES, Packer RJ, Li Y, Billups CA, Smith KS, Jaju A, et al. Efficacy of Carboplatin and Isotretinoin in Children With High-risk Medulloblastoma: A Randomized Clinical Trial From the Children's Oncology Group. JAMA Oncol. 2021;7(9):1313\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar R, Smith KS, Deng M, Terhune C, Robinson GW, Orr BA, et al. Clinical Outcomes and Patient-Matched Molecular Composition of Relapsed Medulloblastoma. J Clin Oncol. 2021;39(7):807\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSavjani R. nnU-Net: Further Automating Biomedical Image Autosegmentation. Radiol Imaging Cancer. 2021;3(1):e209039.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKonar D, Bhattacharyya S, Gandhi TK, Panigrahi BK, Jiang R. 3-D Quantum-Inspired Self-Supervised Tensor Network for Volumetric Segmentation of Medical Images. IEEE Trans Neural Netw Learn Syst. 2023;PP.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchwalbe EC, Lindsey JC, Nakjang S, Crosier S, Smith AJ, Hicks D, et al. Novel molecular subgroups for clinical classification and outcome prediction in childhood medulloblastoma: a cohort study. Lancet Oncol. 2017;18(7):958\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColtin H, Sundaresan L, Smith KS, Skowron P, Massimi L, Eberhart CG, et al. Subgroup and subtype-specific outcomes in adult medulloblastoma. Acta Neuropathol. 2021;142(5):859\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoomro TA, Zheng L, Afifi AJ, Ali A, Soomro S, Yin M, Gao J. Image Segmentation for MR Brain Tumor Detection Using Machine Learning: A Review. IEEE Rev Biomed Eng. 2023;16:70\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEran A, Ozturk A, Aygun N, Izbudak I. Medulloblastoma: atypical CT and MRI findings in children. Pediatr Radiol. 2010;40(7):1254\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDasgupta A, Gupta T, Maitre M, Kalra B, Chatterjee A, Krishnatry R, et al. Prognostic impact of semantic MRI features on survival outcomes in molecularly subtyped medulloblastoma. Strahlenther Onkol. 2022;198(3):291\u0026ndash;303.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Chen C, Fu R, Zhang Y, Fan Y, Xu J, Cen Y. Texture Analysis of T1-Weighted Contrast-Enhanced Magnetic Resonance Imaging Potentially Predicts Outcomes of Patients with Non-Wingless-Type/Non-Sonic Hedgehog Medulloblastoma. World Neurosurg. 2020;137:e27-e33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYan J, Liu L, Wang W, Zhao Y, Li KK, Li K, et al. Radiomic Features From Multi-Parameter MRI Combined With Clinical Parameters Predict Molecular Subgroups in Patients With Medulloblastoma. Front Oncol. 2020;10:558162.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlharbi M, Mobark N, Bashawri Y, Abu Safieh L, Alowayn A, Aljelaify R, et al. Methylation Profiling of Medulloblastoma in a Clinical Setting Permits Sub-classification and Reveals New Outcome Predictions. Front Neurol. 2020;11:167.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMagadza T, Viriri S. Deep Learning for Brain Tumor Segmentation: A Survey of State-of-the-Art. J Imaging. 2021;7(2).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichards BA, Lillicrap TP, Beaudoin P, Bengio Y, Bogacz R, Christensen A, et al. A deep learning framework for neuroscience. Nat Neurosci. 2019;22(11):1761\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJayachandran Preetha C, Meredig H, Brugnara G, Mahmutoglu MA, Foltyn M, Isensee F, et al. Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumour response assessment in neuro-oncology: a multicentre, retrospective cohort study. Lancet Digit Health. 2021;3(12):e784-e94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWaszak SM, Northcott PA, Buchhalter I, Robinson GW, Sutter C, Groebner S, et al. Spectrum and prevalence of genetic predisposition in medulloblastoma: a retrospective genetic study and prospective validation in a clinical trial cohort. Lancet Oncol. 2018;19(6):785\u0026ndash;98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo Z, Xin D, Liao Y, Berry K, Ogurek S, Zhang F, et al. Loss of phosphatase CTDNEP1 potentiates aggressive medulloblastoma by triggering MYC amplification and genomic instability. Nat Commun. 2023;14(1):762.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Medulloblastoma subgroups, MRI, Deep learning, Risk stratification, Convolutional neural network","lastPublishedDoi":"10.21203/rs.3.rs-6622165/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6622165/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjectives\u003c/h2\u003e \u003cp\u003eDeep learning (DL) based on MRI of medulloblastoma enables risk stratification, potentially aiding in therapeutic decisions. This study to develop DL models that identify four medulloblastoma molecular subgroups and prognostic related genetic signatures.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e \u003cp\u003eIn this retrospective study, consecutive patients with newly diagnosed MB at MRI (T1-, T2- and contrast-enhanced T1-weighted) at two medical institutes between January 2015 and June 2023 were identified. Two-stage sequential DL models were designed\u0026mdash;MB-CNN that first identifies wingless (WNT), sonic hedgehog (SHH), Group 3, and Group 4. Further, prognostic related genetic signatures DL models (MB-CNN_TP53/MYC/Chr11) were developed to predict TP53 mutation, MYC amplification, and chromosome 11 loss status. A hybrid model combining MB-CNN and conventional data (clinical information and MRI features) was compared to a logistic regression model constructed only with conventional data. Four-classification tasks were evaluated with confusion matrices (accuracy) and two-classification tasks with ROC curves (area under the curve, AUC).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe datasets comprised 449 patients (mean age\u0026thinsp;\u0026plusmn;\u0026thinsp;SD at diagnosis, 13.55 years\u0026thinsp;\u0026plusmn;\u0026thinsp;2.33, 249 males). MB-CNN accurately classified MB subgroups in the external test dataset (accuracy was in the range of 76.29\u0026ndash;78.71%). MB-CNN_TP53/MYC/Chr11 models effectively predicted signatures (AUC of TP53 in SHH: 0.91, MYC amplification in Group 3: 0.87, chromosome 11 loss in Group 4: 0.89). The accuracy of hybrid model outperformed the logistic regression model (82.20% vs. 59.14%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.009) and showed comparable performance to MB-CNN (82.20% vs. 77.50%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.105).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eMRI-based DL models allowed identification of the molecular medulloblastoma subgroups and prognostic related genetic signatures.\u003c/p\u003e","manuscriptTitle":"Exploring Deep Learning and Hybrid Approaches in Molecular Subgrouping and Prognostic Related Genetic Signatures of Medulloblastoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-03 07:03:07","doi":"10.21203/rs.3.rs-6622165/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"da412981-7251-48f1-852e-b7b31aa57096","owner":[],"postedDate":"June 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-22T16:01:12+00:00","versionOfRecord":{"articleIdentity":"rs-6622165","link":"https://doi.org/10.1186/s41016-025-00405-7","journal":{"identity":"chinese-neurosurgical-journal","isVorOnly":false,"title":"Chinese Neurosurgical Journal"},"publishedOn":"2025-09-15 15:57:33","publishedOnDateReadable":"September 15th, 2025"},"versionCreatedAt":"2025-06-03 07:03:07","video":"","vorDoi":"10.1186/s41016-025-00405-7","vorDoiUrl":"https://doi.org/10.1186/s41016-025-00405-7","workflowStages":[]},"version":"v1","identity":"rs-6622165","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6622165","identity":"rs-6622165","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}