Beating the Curse of Dimensionality: Applying Unbalanced Optimal Transport Clustering of Delta Radiomics to Predict Survival Outcomes in Lung Cancer Patients Treated with Immunotherapy
H Veeraraghavan1*, J Oh2, J Zhu3, J Jiang4, A Tannenbaum5, D Gomez6, J Deasy7, (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) Memorial Sloan Kettering Cancer Center, New York, NY, (3) Stony Brook University, NY, (4) Memorial Sloan Kettering Cancer Center, New York, NY, (5) Stony Brook University, NY,(6) Memorial Sloan Kettering Cancer Center, New York, NY (7) Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
SU-I400-BReP-F2-1 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Generalizable high-dimensional radiomics analyses require methods robust to large numbers of correlated features with varying scales. Hence, we developed an unsupervised Clustering on feature Distributions using unbalanced Optimal mass transport distance (CDO) and demonstrated its use to predict non-small cell lung cancer (NSCLC) immunotherapy outcomes. Materials/
Methods: One hundred and forty-three (100 in discovery, 43 in testing) stage III-IV NSCLC patients treated with immunotherapy and imaged with contrast enhanced CT at baseline and 9 weeks during treatment were retrospectively analyzed. The CERR radiomics toolbox extracted 141 features, including delta radiomics features, from deep learning segmented tumors. First, feature islands of highly correlated features determined using Spearman rank correlation were extracted through clustering. Next, patients were grouped using consensus clustering with unbalanced optimal mass transport (UOMT) distance between patient feature islands. Patients in test set were assigned to appropriate clusters based on UOMT distance to the cluster centers. Correlations with RECIST response (partial/complete vs stable vs progression) using Fisher exact test, and progression free survival (PFS)/overall survival (OS) using uni- and multi-variable Cox regression (adjusting for age, gender, and smoking status), were assessed.
Results: Six feature islands were extracted. One feature island, consisting of 15 different features including 13 delta radiomics features, produced two patient clusters, which predicted response (p = 0.004), PFS (HR 0.507, 95% CI of 0.325 to 0.791, p = 0.003) and OS (HR of 0.548, 95% CI of 0.327 to 0.919, p = 0.023). The test set predicted response (p = 0.010), PFS (HR of 2.15, 95% CI of 1.08 to 4.29, p = 0.029) but not OS (HR of 1.66, 95% CI of 0.786 to 3.52, p = 0.18).
Conclusion: CDO is a promising approach to generate novel imaging biomarkers, with the potential to guide clinicians in selecting patients with metastatic lung cancer for immunotherapy.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email