Non-invasive prediction of Ki-67 and p53 biomarkers in spinal ependymoma via deep learning: using multimodal magnetic resonance imaging and clinical data

Abstract

BACKGROUND: Spinal ependymoma prognosis is closely correlated with tumor malignancy and biomarker levels, such as Ki-67 and p53, which reflect cellular proliferation and genetic instability. Despite their clinical significance, current methods to assess these biomarkers rely on invasive postoperative immunohistochemistry (IHC), delaying critical treatment decisions and limiting preoperative planning. While deep learning have revolutionized biomarker prediction in brain tumors, their application to spinal ependymomas remains underexplored due to the rarity of these tumors, insufficient datasets, and the technical challenges of analyzing spinal cord MRI. We used a deep learning model to predict molecular markers for spinal ependymomas using preoperative magnetic resonance imaging (MRI) scans and clinical information to predict biomarkers for spinal ependymoma. METHODS: This study enrolled 352 patients with preoperative MRI, confirmed histological diagnoses of spinal ependymomas, and Ki-67 and p53 status assessed via IHC. Cross-validation and external testing strategies ensured the generalizability of the results. We harnessed multimodal information by integrating the sagittal and transverse MRI phases with clinical data. MRI scans were automatically segmented to extract high-quality features. These features were used to train an ensemble neural network model, Light Gradient Boosting Machine Net (LGBMNet), which predicted the expression of Ki-67 and p53 biomarkers. To validate model architecture and input choice, we conducted ablation and comparison experiments across multiple classifiers and feature subsets. RESULTS: High-precision automatic image segmentation was achieved using the SegFormer model. LGBMNet showed superior predictive power in cross-validation for Ki-67 and p53, with area under the receiver operating characteristic curves (AUCs) of 0.8904 and 0.8948, and externally validated with AUCs of 0.8348 and 0.8521, respectively. The full multimodal LGBMNet model consistently outperformed reduced and classical variants, highlighting the added value of neural-enhanced fusion. CONCLUSIONS: This study developed a deep learning framework for non-invasive prediction of Ki-67 and p53 in spinal ependymomas, integrating multimodal MRI and clinical data. The SegFormer model achieved high-precision segmentation, ensuring robust feature extraction. LGBMNet, combining Multilayer Perceptron and Light Gradient Boosting Machine, demonstrated strong predictive performance. Our results confirm that deep learning can effectively predict tumor biomarkers preoperatively, aiding precision neurosurgery.

Ähnliche Arbeiten