Click here to


Are you sure ?

Yes, do it No, cancel

Lesion-Level Response Prediction From 18F-FDG PET/CT in Metastatic Melanoma Patients Treated with Immune Checkpoint Blockade

D Huff1*, R Shah2 , A Weisman1 , L Zurbriggen3 , M Albertini3 , R Jeraj1,4 , (1) University of Wisconsin-Madison, Madison, WI, (2) North Memorial Healthcare, Robbinsdale, MN, (3) University of Wisconsin Carbone Cancer Center, Madison, WI, (4) University of Ljubljana, Ljubljana, Slovenia


(Sunday, 7/14/2019) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 5

Purpose: The extraction of quantitative descriptors from lesions seen on molecular imaging can benefit patients receiving immune checkpoint blockade by improving patient stratification and treatment selection. The goal of this work was to identify PET/CT-derived quantitative imaging features capable of predicting lesion-level response to immunotherapy.

Methods: ¹�F-FDG PET/CT scans (N=138) of a cohort of 20 metastatic melanoma patients receiving immune checkpoint inhibitors were analyzed retrospectively. Lesions larger than 1 cm³ (N=198) were identified from clinical reports and segmented using an active-contours method. Fifty-five uptake, shape, and texture features (radiomics) were extracted from each lesion. Lesion response was assigned as either immediate-response (present on one scan only, N=118), eventual-response (present on multiple scans, then resolving, N=52), or non-response (did not resolve, N=28). Univariate analyses assessed differences in feature values between response categories (clustered Wilcoxon rank-sum test). Additionally, supervised classifiers were trained with 5-fold cross-validation to predict lesion immediate-response vs non-immediate-response from extracted features, with area under the curve (AUC) as performance metric.

Results: Univariate analyses showed that SUVtotal of non-responding lesions was significantly higher than that of immediate-responding lesions (p=0.02, Bonferroni corrected). No other univariate baseline feature differed significantly by response category, including tumor volume (p=0.40, Bonferroni corrected). For lesions matched between scans, no longitudinal change in lesion feature was significantly different between non- and eventual-responding lesions. The highest performing machine learning classifier was a Random Forest trained on baseline features selected via LASSO (SUVmax, SUVmean, σSUV, surface area) (AUC=0.77, sensitivity:62%, specificity:85%). For comparison, when used as a univariate predictor of IR vs non-IR, baseline SUVtotal achieved an AUC of 0.72.

Conclusion: Baseline lesion SUVtotal differed significantly between non-responding and immediately-responding lesions. The combination of multiple FDG PET features used as input for supervised machine learning classifiers demonstrated increased performance in predicting lesion response over SUV(total) used as a univariate predictor.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by the University of Wisconsin Carbone Cancer Center Support Grant P30 CA014520, and by the NCI of the NIH under Award Number T32CA009206. The content does not necessarily represent the official views of the NIH. Robert Jeraj is a co-founder of AIQ Solutions.


FDG PET, Quantitative Imaging, Texture Analysis


IM/TH- Image Analysis (Single modality or Multi-modality): Quantitative imaging

Contact Email