PET image nonuniformity texture features for metastasis risk prediction in osteosarcoma.

Abstract

Objective: PET image analysis provides tumor heterogeneity data related to neoadjuvant chemotherapy response (NACR) and metastatic risk in osteosarcoma. Ki-67 expression is used to predict metastasis. The accuracy of prediction models with image quantitative features can be improved by including genetic information. Here, we aimed to evaluate the accuracy of a combination of heterogeneous 18F-fluorodeoxyglucose PET image texture features and Ki-67 expression as predictive indicators of metastasis.

Methods: PET images and clinical data of 82 patients with osteosarcoma before and after treatment were collected. Quantitative features were extracted from the PET images obtained before treatment, and the area under the receiver operating characteristic curve (AUC) for NACR and metastatic event was calculated. Relative risk and odds analyses of the quantitative features of the entire image were performed. Kaplan-Meier survival analysis was performed to determine the relationship between image quantitative features and clinical information. The machine learning prediction model was evaluated using valid image quantitative features and various algorithms of the univariate analysis.

Results: Forty-seven image textures were obtained. The AUC values were 0.504-0.62 for NACR and 0.510-0.598 for metastatic events. The NACR and metastatic risk were related to the gray-level run length matrix (GLRLM) run length nonuniformity (RLNU) (relative risk: 1.3846, P = 0.0138 for NACR; relative risk: 2.1284, P = 0.049 for metastatic event) in the univariate analysis. The accuracy of the prediction model using the random forest algorithm with GLRLM RLNU, Ki-67 expression, and NACR was 0.91 for metastatic risk. NACR and metastatic risk were predicted with high accuracy using the nonuniformity in PET image texture.

Conclusion: Combining PET image texture nonuniformity with Ki-67 expression and clinical data can enhance the accuracy of metastasis prediction in osteosarcoma. This multimodal approach may support metastasis risk prediction in osteosarcoma and aid in personalized treatment planning.