ESTRO 2023

Session Item

Radiomics, modelling and statistical methods

Session Code: 7011

Session Type: Poster (Digital)

Track: Physics

Machine learning prediction of pain response to palliative radiation therapy with CT-based radiomics

Oscar Llorian, Germany

Presentation Number: PO-2106

Abstract

Abstract Title:

Machine learning prediction of pain response to palliative radiation therapy with CT-based radiomics

Authors:

Oscar Llorian^1,2,3, Joachim Akhgar¹, Steffi Pigorsch¹, Kai Borm⁴, Stefan Münch⁴, Denise Bernhardt^4,5,6, Burkhard Rost⁷, Miguel Andrade⁸, Stephanie Combs^4,9,6, Jan Peeken^4,9,6

¹Klinikum rechts der Isar, Technical University of Munich, Department of Radiation Oncology, Munich, Germany; ²School of Computation, Information and Technology, Technical University of Munich, Department for Bioinformatics and Computational Biology, Garching, Germany; ³Institute of Organismic and Molecular Evolution, Johannes Gutenberg University Mainz , Computational Biology and Data Mining, Mainz, Germany; ⁴Klinikum rechts der Isar, Technical University of Munich , Department of Radiation Oncology, Munich, Germany; ⁵Institute of Radiation Medicine, Helmholtz Zentrum, Department of Radiation Sciences, Munich, Germany; ⁶Deutsches Konsortium für Translationale Krebsforschung, Deutsches Konsortium für Translationale Krebsforschung, Munich, Germany; ⁷School of Computation, Information and Technology, Technical University of Munich , Department for Bioinformatics and Computational Biology, Garching, Germany; ⁸Institute of Organismic and Molecular Evolution, Johannes Gutenberg University Mainz, Computational Biology and Data Mining, Mainz, Germany; ⁹Institute of Radiation Medicine, Helmholtz Zentrum , Department of Radiation Sciences, Munich, Germany

Show Affiliations

Purpose or Objective

Painful Spinal Bone Metastases (PSBMs) patients regularly receive palliative Radiation Therapy (RT) with response rates in about 2/3 of patients. In this study, we evaluated the value of Machine Learning (ML) models based on radiomics and semantic imaging features, as well as clinical parameters to predict complete pain response.

Material and Methods

Gross Tumour Volumes (GTV) and Clinical Target Volumes (CTV) of 261 PSBMs were segmented on planning computed tomography (CT) scans. 105 radiomics features were extracted and pre-processed from both volumes of interest. Semantic features from the Spinal Instability Neoplastic Score (SINS), and clinical features, were determined and collected for all patients. ML techniques, including random forest classifier (RFC) and support vector machine (SVM), were trained on the radiomics, semantic and clinical features, and compared using repeated nested cross validation. Models trained on combined features were also evaluated for a possible performance increase. Feature importance was analysed via the mean decrease in impurity for RFC models.

Results

The best radiomics classifier was trained on CTV with an Area Under the Receiver-Operator Curve (AUROC) of 0.62 ± 0.01 (RFC). The semantic model achieved a comparable AUROC of 0.63 ± 0.01 (RFC). This was significantly lower than the clinical ML model (SVM, AUROC: 0.80 ± 0.01) and slightly lower than the Spinal Instability Neoplastic Score (SINS; LR, AUROC: 0.65 ± 0.01). A combined SVM model trained on CTV, SINS, and clinical features did not further improve performance (AUROC: 0,74 ± 0,01). Feature selection frequency and importance analysed on combined models, confirmed that clinical and semantic features were selected more often, achieving high importance scores.

Conclusion

In this exploratory study, we could demonstrate that radiomics and semantic analyses of planning CTs allowed for prediction of therapy response to palliative RT, albeit to a limited extent. The best prediction was possible by applying ML modelling to established clinical parameters, which also improved the performance of the combined models that used them.