1022MO Predicting overall survival of patients with melanoma and NSCLC treated with immunotherapy using AI combining total tumor volume and tumor heterogeneity on CT-Scans

Using total tumor volume and AI tumor heterogeneity from CT scans, it may be possible to predict the prognosis of oncology patients, identify those who could benefit from immunotherapy, and provide valuable guidance for treatment. This retrospective study included 607 patients treated from 2014 to 2019. Baseline CT-scans were collected. All visible lesions were outlined on the largest diameter axial slice by 2 senior radiologists with double reading. The total tumor volume (TTV) was computed by adding estimated lesion volumes. Patients were randomly split into train, test and validation sets with a 56-14-30 ratio stratified by cancer type. Texture features were extracted using either radiomics or a neural network. Two heterogeneity signatures were obtained using PCA followed by random survival forests fit to predict overall survival (OS): radiomics heterogeneity risk (RadH) and AI risk (AiH). These were combined with TTV using a Cox model to establish 2 imaging scores (TTV-RadH, TTV-AiH). Cutoffs were determined using maximally selected rank statistics on the train set. Two populations (low risk-high risk) were evaluated using Kaplan-Meier and Log-Rank tests. Correlations were studied using the Pearson correlation coefficient. In total, 19877 lesions were annotated. The train, test and validation sets contained 339, 85 and 183 patients. Preliminary results were computed on the train and test sets. The cut-off for TTV was 102 cm3. On the test set, TTV-RadH low and high risk groups had median OS of 3.09 and 16.29 months, while TTV-AiH had median OS of 3.09 and 19.04 months. All Log-Rank tests were statistically significant (p<0.001). The correlation was strong between TTV and RadH (⍴=0.67), moderate between AiH and RadH (⍴=0.40), and weak between TTV and AiH (⍴=0.27). This study developed new prognostic imaging signatures based on baseline CT-scans. These combine both TTV and heterogeneity indexes, extracted using radiomics or a novel method based on self-supervised learning. These methods were able to predict survival of cancer patients treated with immunotherapy in both train and test sets. RadH was more correlated than AiH with volume.