Multiscale Fusion Models With Genomic, Topological, and Pathomic Features to Predict Response to Radiation Therapy
Yu Jin1, Hidetaka Arimura2, Takeshi Iwasaki3
1Division of Medical Quantum Science, Department of Health Sciences, Graduate School of Medical Sciences, Kyushu University, Fukuoka, Japan.
Abstract:
Artificial intelligence models with biomarkers to predict treatment responses to radiation would be necessary to maximize the treatment outcomes of individual patients, especially with histopathology images routinely obtained before treatment. We hypothesized that multiscale features, such as genomic (GM), pathomic (PM), and topological (TP) features, could be associated with the radiation response. We investigated fusion models with multiscale features in histopathology images to predict response to radiation therapy for patients (responders) with non-small cell lung cancer. Ten radiosensitivity-related (radiosensitive and radioresistant) genes were deployed as GM features. PM features were extracted from histopathology images by conventional PM analyses. TP features represent the intrinsic properties of tumor cells using Betti numbers, which are mathematical invariants. We analyzed non-small cell lung cancer patients from The Cancer Genome Atlas and Clinical Proteomic Tumor Analysis Consortium who received radiotherapy and established 3 base models with GM, TP, and PM features, respectively, and 3 fusion models. The TP model showed a higher area under the receiver operating characteristic curve of 0.707 (P = .026, log-rank test in overall survival analysis) in the internal test data set and 0.720 (P = .136) in the external test data set. The results indicated that the TP models achieved better classification and prognostic prediction powers than the other base models. The inner-cell TP structure may have the ability to reveal the cell radiosensitivity-related information. Furthermore, the best fusion model with GM, TP, and PM features achieved the highest area under the receiver operating characteristic curve of 0.846 (P = .019) and 0.731 (P = .043) in predicting the treatment response and prognoses in the internal and external test data sets, respectively. This study demonstrated the predictive power of the multiscale fusion model for histopathology images, which may assist clinical physicians in the selection of responders to radiation for personalized radiation therapy and would be substantially beneficial for patients with cancer.
