Application Of Deep Learning For Clinical Predictive Modeling: An Artificial Intelligence Recognition In Spinal Metastases.

e18050 Background: Spinal metastases are very common outcomes within solid malignant tumors, which could lead to various skeletal related events (SREs). The accurate and timely diagnosis is the key to improve prognosis. Recently, artificial intelligence(AI) has assisted doctors in many ways by different AI technologies. In this study, we applicated a deep learning model to classify and locate the metastatic lesions on spinal CT images. Methods: We set up a dataset consisting of 800 patients’ spinal CT images, which contained over 300,000 CT slices. And we built a multi-label classification and vertebrae segmentation model to recognize the metastatic lesions on spinal CT images. Then we trained and tested this model within our dataset, using a data augmentation by random flips and random rotations. Sensitivity and specificity were used to evaluate the performance of the model. Results: Our model showed that the diagnostic utilities of normal lesions were: sensitivity 81.7% and specificity 92%; while the diagnostic utilities of metastatic lesions were: sensitivity 84.7% and specificity 84.5%. Conclusions: Our model can effectively and accurately discriminate spinal metastases on spinal CT images. [Table: see text]