P13.17.B PREDICTING KI-67 PROLIFERATION INDEX OF MENINGIOMAS ON MRI: A DEEP LEARNING METHOD BASED ON MULTI-MODAL INFORMATION

Abstract BACKGROUND A high Ki-67 index usually suggests accelerated cell proliferation of meningioma related to significant tumor growth as well as increased recurrent risk. This study aimed to explore the feasibility of deep learning method in predicting Ki-67 index of meningiomas with multi-modal information. MATERIAL AND METHODS Pre-treatment magnetic resonance images were retrospectively curated from 521 patients with surgically resected, pathologically confirmed meningiomas from three institutions. The cases were classified into low-expressed or high-expressed groups with a threshold of 5% of Ki-67 index. Predictive models were developed with multi-modal deep learning network by using traditional radiological findings, radiomics features extracted from tumors, and MRIs of meningiomas. The performance of the models was evaluated with area under curve (AUC), accuracy (ACC), sensitivity, and specificity. In addition, 127 cases with incidental small meningioma were recruited and followed up in 2 years, to investigate if the model could be used for predicting the tumor growth to assist in patient management. RESULTS Overall, 371 patients were enrolled for model development and primary analysis. The predictive model showed good performance with AUC of 0.798, ACC of 0.710, sensitivity of 0.613, and specificity of 0.806 in the internal test. It also achieved robustness in the external test cohort consisted of 150 cases, with AUC of 0.758, ACC of 0.661, sensitivity of 0.677, and specificity of 0.645. Moreover, model-predicted high Ki-67 tumor was associated with significant tumor volume growth happened in two years. CONCLUSION The predictive model can efficiently predict the Ki-67 index in meningioma patients, and showed good potential in facilitating the therapeutic decisions.