ESTRO 2024 - Abstract Book

S5137

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Patients’ cohort consisted in 18 H&N patient treated with a prescription dose of 70 Gy using dual-arc VMAT were included. Two different patients’ sub-cohort were identified: 9 patients with mandible’s osteoradionecrosis (ORN) and 9 patients as control dataset, without ORN. The delivered plan for each patient was calculated in the Eclipse treatment planning system with the Analytical Anisotropic Algorithm v13.6 (AAA), that yielded dose to water-in-water (Dw,w), and with Acuros XB v13.6 (AXB) for both dose to medium-in-medium (Dm,m) –AXB Dm–, and dose to water-in-medium (Dw,m) –AXB Dw– reporting quantities (figure 1). The dose distributions presented a similar pattern in each algorithm: for AAA -for which the plan was originally optimized- the dose distribution within the PTV was homogeneous; and for AXB Dm, the dose values in the mandible were lower while for AXB Dw were higher.

Figure 1 – Example of dose distribution calculated by AXB Dw, AAA and AXB Dm (modified based on [2]). Different “hot spot” is highlighted in the PTV containing mandibular bone. Maximum dose is 82.3 Gy, 76.8 Gy and 73.3 Gy for AXB Dw, AAA and AXB Dm respectively. An automated software (iTA Mining, Tecnologie Avanzate, Turin, Italy) for the multi-parameter’s radiotherapy model prediction was employed for the data import, quality check, dosiomic feature extraction, features redundancy process, and model training and evaluation. Dosiomic features were extracted from the mandible structure with MODDICOM (open-source R library, IBSI compliant) [1]. The following features were calculated considering two aggregation methods (2D and 2.5D): 17 intensity-based statistics (STAT), 100 grey level co-occurrence matrix (GLCM), 65 grey level run length matrix (GLRLM) and 32 grey level size zone matrix (GLSZM). Then a correlation-based feature selection was applied to identify the most significant ones for each dose calculation algorithm and dose reporting mode. A Support Vector Machine (SVM) with five-fold cross-validation was used for the mandibular toxicity prediction model.

The performance was evaluated in terms of specificity, sensibility, precision and accuracy.

Results:

Dosiomic features reduction lead to the following sets of relevant features: STAT Robust Mean Absolute Deviation (stat.rmad), GLSZM Small zone high gray-level emphasis (szm.szhge) and GLCM Energy with 2.5D aggregation (cm_2.5D.energy) for AAA plans; stat.rmad, cm_2.5D.energy, GLRLM Run percentage (rlm.r.perc), and GLSZM Small zone high gray-level emphasis with 2.5D aggregation (szm_2.5D.szhge) for AXBDw; stat.rmad, cm_2.5D.energy, GLCM Joint maximum with 2.5D aggregation (cm_2.5.joint.max) and STAT 90th percentile (stat.90thpercentile) for AXBDm. The common most relevant dosiomic features were stat.rmad and cm_2.5D.energy. With these sets of features SVM reached an equivalent training accuracy of 94%. A new training was run considering the two features stat.rmad and cm_2.5D.energy that were relevant for all three dose calculations. Table 1 shows the model performance predicting ORN for the three dose calculation algorithms. The two dosiomic features (stat.rmad and cm_2.5D.energy) seems to