Lavinia Monaco1, Elisabetta De Bernardi2,3, Francesca Bono4, Diego Cortinovis5, Cinzia Crivellaro6, Federica Elisei6, Vincenzo L'Imperio4, Claudio Landoni2,6, Gregory Mathoux2, Monica Musarra6, Fabio Pagni4, Elia Anna Turolla2,7, Cristina Messa2, Luca Guerra2,6. 1. School of Medicine and Surgery, University of Milan-Bicocca, Monza, Lombardy, Italy. laviniamonaco92@gmail.com. 2. School of Medicine and Surgery, University of Milan-Bicocca, Monza, Lombardy, Italy. 3. Centro di Ricerca Interdipartimentale Bicocca Bioinformatics Biostatistics and Bioimaging Centre - B4, Università Degli Studi Di Milano-Bicocca, Milan, Italy. 4. Department of Medicine and Surgery, Pathology, ASST Monza, San Gerardo Hospital, University of Milan-Bicocca, Monza, Italy. 5. Medical Oncology, ASST Monza, San Gerardo Hospital, Monza, Italy. 6. Nuclear Medicine Department, ASST Monza, San Gerardo Hospital, Monza, Lombardy, Italy. 7. Tecnomed Foundation, University of Milan-Bicocca, Milan, Italy.
Abstract
PURPOSE: The present pilot study investigates the putative role of radiomics from [18F]FDG PET/CT scans to predict PD-L1 expression status in non-small cell lung cancer (NSCLC) patients. METHODS: In a retrospective cohort of 265 patients with biopsy-proven NSCLC, 86 with available PD-L1 immunohistochemical (IHC) assessment and [18F]FDG PET/CT scans have been selected to find putative metabolic markers that predict PD-L1 status (< 1%, 1-49%, and ≥ 50% as per tumor proportion score, clone 22C3). Metabolic parameters have been extracted from three different PET/CT scanners (Discovery 600, Discovery IQ, and Discovery MI) and radiomics features were computed with IBSI compliant algorithms on the original image and on images filtered with LLL and HHH coif1 wavelet, obtaining 527 features per tumor. Univariate and multivariate analysis have been performed to compare PD-L1 expression status and selected radiomic features. RESULTS: Of the 86 analyzed cases, 46 (53%) were negative for PD-L1 IHC, 13 (15%) showed low PD-L1 expression (1-49%), and 27 (31%) were strong expressors (≥ 50%). Maximum standardized uptake value (SUVmax) demonstrated a significant ability to discriminate strong expressor cases at univariate analysis (p = 0.032), but failed to discriminate PD-L1 positive patients (PD-L1 ≥ 1%). Three radiomics features appeared the ablest to discriminate strong expressors: (1) a feature representing the average high frequency lesion content in a spherical VOI (p = 0.009); (2) a feature assessing the correlation between adjacent voxels on the high frequency lesion content (p = 0.004); (3) a feature that emphasizes the presence of small zones with similar grey levels inside the lesion (p = 0.003). The tri-variate linear discriminant model combining the three features achieved a sensitivity of 81% and a specificity of 82% in the test. The ability of radiomics to predict PD-L1 positive patients was instead scarce. CONCLUSIONS: Our data indicate a possible role of the [18F]FDG PET radiomics in predicting strong PD-L1 expression; these preliminary data need to be confirmed on larger or single-scanner series.
PURPOSE: The present pilot study investigates the putative role of radiomics from [18F]FDG PET/CT scans to predict PD-L1 expression status in non-small cell lung cancer (NSCLC) patients. METHODS: In a retrospective cohort of 265 patients with biopsy-proven NSCLC, 86 with available PD-L1 immunohistochemical (IHC) assessment and [18F]FDG PET/CT scans have been selected to find putative metabolic markers that predict PD-L1 status (< 1%, 1-49%, and ≥ 50% as per tumor proportion score, clone 22C3). Metabolic parameters have been extracted from three different PET/CT scanners (Discovery 600, Discovery IQ, and Discovery MI) and radiomics features were computed with IBSI compliant algorithms on the original image and on images filtered with LLL and HHH coif1 wavelet, obtaining 527 features per tumor. Univariate and multivariate analysis have been performed to compare PD-L1 expression status and selected radiomic features. RESULTS: Of the 86 analyzed cases, 46 (53%) were negative for PD-L1 IHC, 13 (15%) showed low PD-L1 expression (1-49%), and 27 (31%) were strong expressors (≥ 50%). Maximum standardized uptake value (SUVmax) demonstrated a significant ability to discriminate strong expressor cases at univariate analysis (p = 0.032), but failed to discriminate PD-L1 positive patients (PD-L1 ≥ 1%). Three radiomics features appeared the ablest to discriminate strong expressors: (1) a feature representing the average high frequency lesion content in a spherical VOI (p = 0.009); (2) a feature assessing the correlation between adjacent voxels on the high frequency lesion content (p = 0.004); (3) a feature that emphasizes the presence of small zones with similar grey levels inside the lesion (p = 0.003). The tri-variate linear discriminant model combining the three features achieved a sensitivity of 81% and a specificity of 82% in the test. The ability of radiomics to predict PD-L1 positive patients was instead scarce. CONCLUSIONS: Our data indicate a possible role of the [18F]FDG PET radiomics in predicting strong PD-L1 expression; these preliminary data need to be confirmed on larger or single-scanner series.