Abstract
1180
Objectives: Deep neural network (DNN) based artificial intelligence model has indicated considerable efficiency and accuracy in diagnosis of normal and diffused cancer bone metastases according to patients, but still falls short in identifing few-lesions (lesion number ≤5) images. We aim to explore a new AI model based on DNN and evaluate its diagnostic capability of individual lesion by the distribution characteristics in lung, breast and prostate cancers.
Methods: The study included 9461 lesions in 2337 patients with lung, breast or prostate cancer, who had received 99mTc-MDP bone imaging for AI training. According to the site of metastasis, the lesions were divided into five groups by body area: skull, sternum, vertebrae, pelvis and limb bone. Then, the distribution and differential features of each lesion in these body areas were labeled by accompanying CT or MR information, and used for AI learning. After training, another testing cohort containing 587 patients was selected for evaluating the diagnostic performance of AI model. The sensitivity, specificity, accuracy, positive predictive value, negative predictive value, receiver operating characteristic (ROC) and areas under the curve (AUC) of lesion identification for each cancer were compared.
Results: The accuracy of the artificial intelligence model for overall and focus-based diagnosis of the three tumors was significantly higher in prostate cancer (84.26%/82.82%) than in lung cancer (76.42%/74.30%) and breast cancer (79.78%/76.30%). The AUC values of prostate, lung, and breast cancer were 0.879, 0.794 and 0.838. The AUC values for the diagnostic performance of each part of the model are skull: 0.720, sternum: 0.776, vertebra: 0.891, pelvis: 0.801 and limb bone: 0.919. The distribution of metastatic lesions of lung cancer and breast cancer in the sternum and vertebra was significantly different from that in other parts, but there was no significant difference between them. In prostate cancer, the proportion of metastatic lesions in the pelvis was significantly higher than that in other parts of the pelvis.
Conclusions: The distribution of bone metastases in lung cancer, breast cancer and prostate cancer has its own characteristics. The artificial intelligence model based on deep neural network has great potential in helping nuclear medicine doctors to diagnose tumor bone metastasis, but its performance needs to be improved and more auxiliary diagnostic information needs to be included, such as blood biochemistry, CT, MRI test results, and even its treatment information.