1454P - Radiomic analysis predicts response to immunotherapy in metastastic non-small cell lung cancer (mNSCLC): Preliminary results

Background

Immunotherapy (IT) with single-agent immune-checkpoint inhibitors (ICI) is standard treatment in all comers mNSCLC patients (pts) with PD-L1 expression >50%. However, inter- and intratumoral heterogeneity of PD-L1 expression limits its predictive value. Radiomics has the potential to integrate imaging and biological data thus recapitulating tumor tissue heterogeneity. We used a machine learning approach to test computed tomography (CT) tumor radiomic features in terms of predictive value of immunotherapy response.

Methods

We retrospectively analyzed radiological and clinical data of 190 pts who received first-line immunotherapy with Pembrolizumab (P) for non-oncogene addicted mNSCLC with PDL1>50% treated at our Institution from 2017 to 2022. Baseline contrast-enhanced CT (CECT) imaging of the primary tumor was contoured to isolate a polygonal region of interest (ROI) enclosing the whole tumor volume and the peripheral tumor area (3mm thickness). After thresholding and filtration of images, radiomic analysis was conducted to isolate 107 radiomic features of mean grey level (GL) values within the ROIs. Median progression-free survival (PFS) was calculated from start of IT to first evidence of progression or death. A machine learning approach with cross-validation (Orange Data Mining Software) was applied to select the best radiomic features in terms of PFS predictive value.

Results

After pre-processing of CECT imaging, 78 eligible pts were finally included in the radiomic analysis. Median PFS was 5 months. In the whole series, 33% and 22% of pts were classified as hyper-responders (PFS>24 months) and hyper-progressive (PFS<1 month), respectively. PDL1 expression level (50-75% vs >75%) was not predictive of PFS at univariate analysis (3 vs 5 months, p 0.147). After machine learning selection with 10-fold cross validation, 30 out of total 107 radiomic features were extracted to best predict PFS (9 vs 2 months, p <0.005).

Conclusions

Preliminary results of this exploratory analysis indicate that as predictive PFS factors, radiomic features outperform common clinico-pathological factors including PD-L1 expression even at thresholds >75% and >90%. Validation of these results is warranted.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.