For a retrospective study involving 607 oncology patients treated from 2014-2019, the potential for predicting patient prognosis and guiding immunotherapy treatment through the analysis of total tumor volume (TTV) and artificial intelligence (AI) tumor heterogeneity from CT scans was investigated. Baseline CT scans were collected, and lesions were outlined on the largest diameter axial slice by two senior radiologists. TTV was calculated by summing estimated lesion volumes. The patients were randomly divided into train, test, and validation sets, maintaining a 56-14-30 ratio stratified by cancer type. Texture features were extracted using radiomics or a neural network, and two heterogeneity signatures—Radiomics Heterogeneity Risk (RadH) and AI Risk (AiH)—were obtained using principal component analysis (PCA) followed by random survival forests. The study established two imaging scores, TTV-RadH and TTV-AiH, by combining TTV with the heterogeneity signatures using a Cox model. Cutoffs were determined on the train set. In total, 19,877 lesions were annotated, with the train, test, and validation sets containing 339, 85, and 183 patients, respectively. Preliminary results on the test set revealed statistically significant differences in median overall survival (OS) between low and high-risk groups for both TTV-RadH (3.09 vs. 16.29 months) and TTV-AiH (3.09 vs. 19.04 months), with p-values below 0.001 in all cases. Strong correlations were observed between TTV and RadH (⍴=0.67), moderate between AiH and RadH (⍴=0.40), and weak between TTV and AiH (⍴=0.27). The study successfully developed novel prognostic imaging signatures using baseline CT scans, integrating TTV and heterogeneity indexes derived from radiomics or a novel self-supervised learning method. These imaging scores demonstrated the ability to predict the survival of cancer patients treated with immunotherapy in both the train and test sets, with RadH exhibiting a stronger correlation with volume compared to AiH.
