Computational Prediction of Symptomatic RP Based On Planning Computed Tomography Images Using Topologically Invariant Betti Numbers Prior to Stereotactic Body Radiation Therapy
K Ninomiya1*, H Arimura1, T Yoshitake2, T Hirose2, Y Shioyama3, (1) Kyushu University, Fukuoka, JP (2) Kyushu University Hospital, Fukuoka, JP (3) SAGA HIMAT Foundation, Saga, JP
Presentations
WE-F-TRACK 6-7 (Wednesday, 7/28/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Purpose: Stereotactic body radiation therapy (SBRT) is commonly administrated for inoperable patients with early stage non-small-cell lung cancer. However, some patients, whose lung tissue is vulnerable to radiation, could experience radiation induced pneumonitis (RP). Therefore, we have developed a symptomatic RP prediction model by characterizing the vulnerability of lung regions based on planning computed tomography (pCT) images using topologically invariant Betti numbers (BNs) prior to SBRT.
Methods: Two-hundred sixty-six patients who underwent SBRT were selected for this study. A total of 236 (20 with ≥ grade 2 symptomatic RP) and 30 cases (8 with ≥ grade 2 symptomatic RP) were used for training and independent test of the RP prediction models, respectively. A lung region on an axial plane of the pCT image, which included a centroid of a gross tumor volume, was analyzed for characterizing the vulnerability of lung regions. A total of 47,616 image features were obtained from 512 BN maps that may represent topologically invariant heterogeneous characteristics of lung cancer by applying histogram- and texture-based feature calculations. A BN-based signature was constructed using the least absolute shrinkage and selection operator logistic regression model. Support vector machine (SVM) models were utilized to build the RP prediction model using the BN-based signature in a leave-one-out cross validation. The performance of the BN-based model was compared to the SVM model trained with the conventional signature consisting of original pCT image features (pCT-based model) regarding a robustness index (RI) which defined higher total and lower difference of area under receiver operating characteristics curves (AUCs) in the validation and test.
Results: The RIs for BN- and pCT-based models were 1.544 (AUC in the validation: 0.800, test: 0.778) and 1.191 (AUC in the validation: 0.750, test: 0.608), respectively.
Conclusion: The BN-based model could be feasible for prediction of symptomatic RP prior to SBRT.
Handouts
Keywords
Treatment Planning, Stereotactic Radiosurgery, Lung
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Quantitative imaging
Contact Email