ESTRO 2023 - Abstract Book

S1913

Digital Posters

ESTRO 2023

specimens were placed inside their shape-preserving moulds for ex-vivo T2w imaging, followed by histopathological preparation and subsequent evaluation. The evaluation resulted in digital annotations showing the location of lesions and their International Society of Urological Pathology (ISUP) grade groups (IGG). The in-vivo T2w images served as a common frame of reference for all image data. The histopathology was first registered to the ex-vivo T2w, which in turn was co- registered to the in-vivo T2w. We used PET and mpMRI data to differentiate between ISUP grades in two cases: IGG 3 vs. IGG 2 and IGG ≥ 3 vs. IGG ≤ 2. PET uptake was quantified in maximum standardized uptake values (SUV max ). mpMRI was quantified in relative measures of the maximum volume transfer constant ( R K trans ), median apparent diffusion coefficient ( R ADC) and the median intensity of T2w images ( R T2w). The superscript R signifies that these values were normalized to the corresponding mean intensities of healthy prostatic tissue. The performance was evaluated by receiver operating characteristics (ROC) analysis, using a fast implementation of Delong’s algorithm to compare the area under the ROC curves (AUCs). Results We identified 660 lesions in total. The results are based on the 178 lesions having an in-plane area larger than a circle with a radius roughly equal to the upper limit of the estimated uncertainty of the registration method ( ≥ 20 mm ² ). Cut-off values for PSMA-SUV max combined with Acetate-SUV max , R K trans and R ADC yielded AUCs of 0.88 for IGG 3 vs. IGG 2 and 0.89 for IGG ≥ 3 vs. IGG ≤ 2. PSMA-SUV max alone achieved an AUC of 0.73 for IGG 3 vs. IGG 2, and 0.78 for IGG ≥ 3 vs. IGG ≤ 2. The corresponding figures for biparametric MRI ( R ADC and R T2w) were 0.63 and 0.59. mpMRI ( R K trans , R ADC and R T2w) increased the AUCs to 0.74 for IGG 3 vs. IGG 2 and 0.75 for IGG ≥ 3 vs. IGG ≤ 2 (p < 0.01). Combining PSMA-SUV max with mpMRI further increased the AUCs to 0.82 (p < 0.05) and 0.84 (p < 0.01), respectively. See Fig. 1 for a selection of ROC curves. Conclusion The clinically important distinction between ISUP grade groups 2 and 3 could be reflected in partially discriminative cut- off values derived from PSMA-PET, Acetate-PET and mpMRI. The results also indicate that the imaging methods provide independent information. However, ADC-maps and T2w images contributed less to the differentiation between ISUP grade groups.

PO-2125 Contrastive self-supervised learning of lung tumor imaging predicts immunotherapy response

T. Chaunzwa 1,2 , S. Pai 2 , D. Bontempi 2 , B. Ricciuti 3 , S. Bernatz 2 , J. Alessi 3 , R. Mak 1,2 , M. Awad 3 , H. Aerts 1,2

1 Dana Farber Brigham Cancer Center, Harvard Medical School, Radiation Oncology, Boston, USA; 2 Mass General Brigham, Harvard Medical School, Artificial Intelligence in Medicine Program, Boston, USA; 3 Dana Farber Brigham Cancer Center, Harvard Medical School, Medical Oncology, Boston, USA Purpose or Objective Chemo-immunotherapy improves survival only in a subset of advanced non-small cell lung cancer (NSCLC) patients, and established response biomarkers, such as PD-L1 expression, have limited predictive value. In this study, we aim to develop a robust lung cancer imaging biomarker for chemo-immunotherapy response. Materials and Methods A cohort of 209 patients receiving first-line chemo-immunotherapy for advanced NSCLC at Dana-Farber Cancer Institute between 2015 and 2021 was utilized. Baseline thoracic computed tomography (CT) scans obtained prior to initiation of combined therapy with pembrolizumab, carboplatin, and paclitaxel were retrospectively collected and analyzed. Patients without baseline imaging or who have received prior treatment were excluded. A contrastive self-supervised learning (SSL) algorithm pre-trained for visual recognition on more than 11,000 CT data samples (with and without pulmonary lesions) was used for feature extraction. LASSO-Cox regression was used to select features strongly associated with the primary outcome of interest, progression-free survival (PFS). A k-nearest neighbors classifier was then applied to the resultant feature vector as well as a holistic model including additional combinations of clinical, genomic, and immunophenotypic parameters. These classical variables, including PD-L1 expression pattern, tumor mutation burden (TMB), driver mutation status, histology, age, ECOG performance status, BMI, sex, and race, were also independently evaluated. All models were