Purpose: Complex target shape and unique location pose big challenges of plan quality and achievable dosimetric parameters and subsequent treatment outcome for Malignant pleural mesothelioma (MPM) on individual basis. Correlations between target characteristics, planning dosimetric parameters and overall survival were explored using machine learning methods.
Methods: A retrospective cohort of 59 MPM patients, receiving 45 Gy/25 fractions on a Tomotherapy unit, were pooled and analyzed. The achieved dosimetric parameters for target and critical OARs were analyzed for 23 left-sided mesothelioma (LSM) and 36 right-sided mesothelioma patients (RSM) respectively. LASSO function 1/2 ‖Ax-b‖₂²+γ‖x‖₁ was adopted for selecting the most relevant predictors with dosimetric endpoints and/or overall patient survival. Machine learning methods including Multiple Layer Perceptron (MLP) were applied to construct predictive models for MPM response after a pleurectomy/decortication. The model performance was evaluated using the Area Under the ROC Curve (AUC) with 3-fold cross validation.
Results: No differences were observed between planning target volume (PTV) coverage 94.2±1.2% (LSM) vs 93.4±2.4% (RSM). The achieved dosimetric endpoints were found significantly different (p<0.05) for OARs between the LSM and RSM: ipsilateral lung V20 (91.2±6.7 vs. 85.3±6.5%), ipsilateral mean lung doses (39.1±3.8 vs 36.9±2.5 Gy), heart mean doses (25.2 ±3.6 vs 18.8±3.7 Gy), liver mean doses (10.6±2.7 vs. 23.4±4.2 Gy), respectively. There was no significant difference in total lung V20-PTV mean dose 13.4±2.0 vs. 14.4±1.5 Gy nor expected lung pneumonitis for LSM and RSM. The MLP reached an AUC of 0.71 in predicting overall survival from combined characteristics and planning dosimetric parameters. The identified predictors that are most associated with overall survival were age, PTV volume, ipsilateral lung volume and V20Gy, and ipsilateral lung-PTV V20.
Conclusion: Significantly dosimetric differences of OARs were observed for left- and right-side MPM. The predictive models can assist in achieving patient-specific optimal plan quantity and treatment outcome via personalized treatment.