MRI Based Deep Learning Combined Clinical and Imaging Features to Differentiate Medulloblastoma and Ependymoma in Children

Yasen Yimit ¹ Parhat Yasin ² Yue Hao ¹ Abudouresuli Tuersun ¹ Chencui Huang ³ Jingru Wang ³ Xiaoguang Zou ¹ Ya Qiu ¹ Yunling Wang ⁴ Mayidili Nijiati ^5*

• ¹ The First People’s Hospital of Kashi Prefecture, Kashi(Kashgar), China
• ² Sixth Affiliated Hospital of Xinjiang Medical University, Ürümqi, Xinjiang Uyghur Region, China
• ³ Deepwise AI Lab, Beijing, China
• ⁴ First Affiliated Hospital of Xinjiang Medical University, Urumqi, Xinjiang Uyghur Region, China
• ⁵ Fourth Affiliated Hospital, Xinjiang Medical University, Urumqi, China

Background: Medulloblastoma (MB) and Ependymoma (EM) in children share similarities in age group, tumor location, and clinical presentation. Distinguishing between them through clinical diagnosis is challenging. Purpose: This study aims to explore the effectiveness of T2WI-based deep learning combined with clinical imaging features to differentiate MB and EM. Methods: Axial T2-weighted MRI sequences from 201 patients across three centers were utilized for model training and testing. Regions of interest (ROIs) were manually delineated by an experienced neuroradiologist, with oversight from a senior radiologist. We developed a deep learning (DL) classifier using a pretrained AlexNet architecture, which was fine-tuned on our dataset. To mitigate class imbalance, we implemented data augmentation techniques and employed K-fold cross-validation to enhance model generalizability. For patient classification, we utilized two voting strategies: the hard voting strategy, which selected the majority prediction from individual image slices, and the soft voting strategy, which averaged prediction scores across slices with a threshold of 0.5. Additionally, a multimodality fusion model was constructed, integrating the DL classifier with clinical and imaging features. Model performance was assessed using a 7:3 random split for training and validation. Key metrics, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, area under the ROC curve (AUC), and accuracy, were calculated, with statistical comparisons performed using the DeLong test. Medulloblastoma (MB) was classified as positive and ependymoma (EM) as negative. Results: The deep learning model utilizing a hard voting strategy achieved an AUC of 0.712 (95% CI: 0.625-0.797) on the training set and 0.689 (95% CI: 0.554-0.826) on the test set. In contrast, the multi-modality fusion model demonstrated superior performance with an AUC of 0.987 (95% CI: 0.974-0.996) on the training set and 0.889 (95% CI: 0.803-0.949) on the test set. The DeLong test indicated a statistically significant improvement in AUC for the fusion model compared to the DL model (p < 0.001), highlighting its enhanced discriminative ability. Conclusion: T2WI-based deep learning combined with multimodal clinical and imaging features could effectively differentiate MB and EM. The decision tree structure in the decision tree classifier could greatly assist clinicians in daily clinical practice.