A Combined Radiomic and Clinical Model for the Differential Diagnosis of Pneumonitis in Patients with NSCLC (Non-Small Cell Lung Cancer) Patients Receiving Immunotherapy (IO Therapy)
A Traverso1*, F Tohidinezhad1, D Bontempi1, A Dekker1, L Hendriks2, D De Ruysscher1, (1) MAASTRO Clinic, Maastricht, NL (2) Maastricht University Medical Centre, Maastricht, NL
Presentations
SU-H330-IePD-F9-4 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 9
Purpose: IO-therapy induced pneumonitis(IIP) is a rare side effect, but it can degrade patients’ quality of life and determine IO-therapy interruption. The diagnostic challenge is the overlapping clinical manifestations of IIP and Other types of Pneumonitis(OP). A wrong diagnosis leads to over-treatment (steroids) or under-treatment (IO-therapy interruption), decreasing the effectiveness of clinical trials. We developed a combined model to support a better differential diagnosis.
Methods: A clinical trial (NCT03305380) recruited 626 stage IV NSCLC patients receiving anti-PD(L)1 medications as the I/II line treatment from six centers in NL/BE.Radiomic features were extracted from Computed Tomography (CT) images at the time of dyspnea from: A) the segmented lungs (auto-contouring with deep learning), and B) spherical/cubical regions surrounding the center of the highest inflamed regions indicated by the radiologists (semi-automated segmentation). The features were used together with clinical factors to build three models: clinical-only, radiomic-only, and combined. ROC analysis was performed using bootstrap for optimism-corrected results.To evaluate the clinical gain in using the model, the Decision Curve Analysis(DCA) was performed.
Results: A total of n=31 IIP and n=42 events were detected. The clinical model included the following variables: presence of cardiopulmonary commodities (yes/no) and line of IO-therapy treatment. The radiomic model included three wavelet texture features measuring grey-level heterogeneity in the frequency space: "zone entropy", "small dependence low grey-level emphasis", and "dependence entropy". The final model included one clinical factor and two radiomic features. The combined model reached an AUC of 0.75, which was higher than the clinical model. DCAs showed a net benefit for a specific category of low-risk patients. Repeated analysis with the semi-automated approach led to better results (combined model AUC=0.81) and overall benefit in the DCA, showing the importance of introducing clinical knowledge in model design.
Conclusion: Our proof-of-concept combined model can potentially improve treatment in patients receiving IO-therapy.
Keywords
Image Analysis, CT, Computer Vision
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email