Abstract
241952
Introduction: In oncology, most radiomic studies focus on tumor lesions only, ignoring potentially useful signal present in non-tumor regions. However, recent studies suggested that the metabolic activity outside the lesion could include valuable information for predicting overall survival (OS) in cancer patients. We investigated whether the combination of clinical and radiomic features from whole-body [18F]FDG-PET images, including malignant and tumor-free regions, could predict OS in metastatic triple negative breast cancer (TNBC) patients.
Methods: Baseline whole-body [18F]FDG-PET/CT images acquired before any treatment and clinical data were collected for 122 metastatic TNBC patients with at least 24 months of follow-up. All tumor lesions were automatically segmented in the PET images using the AI-driven automated segmentation tool LION (). Lesion delineation was then refined using a standardized uptake value (SUV) threshold of 4. The resulting regions were used to calculate the Total Metabolic Tumor Volume (TMTV) and the distance between the 2 most distant lesions (Dmax) with LIFEx v.7.4.2 [Nioche et al. Cancer Res. 2018]. Using MOOSE v.2.3.5 [Shiyam Sundar et al. J Nucl Med. 2022] and TotalSegmentator v.2.0.5 [Wasserthal et al. Radiology: Artificial Intelligence. 2023], six anatomical structures (subcutaneous fat, muscles, pancreas, spinal cord, spleen, and thyroid) were automatically segmented on the CT of the PET/CT scan. The mean SUV normalized by lean body mass (SULmean) [Sugawara, et al. Radiology, 1999] was computed for each anatomical compartment, except the subcutaneous fat. We also calculated the mean Hounsfield Units (HUmean) and the volumes for the subcutaneous fat and the muscles. All 11 radiomic feature values, as well as the patient age and body mass index (BMI), were dichotomized using cut-off values that maximized the log-rank test statistic for OS. Using a stepwise algorithm with the Akaike Information Criteria (AIC), we identified a Cox multivariable model for OS based on clinical features, TMTV and Dmax (called M1) or based on clinical variables and all 11 radiomic features extracted from non-/malignant regions (M2). For each model, the patients were stratified into 3 groups that led to the highest Harrell’s Concordance-index (C-index). The 1-year, 2-year and 3-year OS rates for each survival profile were reported.
Results: The multivariable model M1 (C-index = 0.73) included the 4 proposed features: older patients (> 45y), with a high dissemination of lesions (Dmax > 56cm), high TMTV (> 167 mL) and low BMI (< 22.6 kg.m-2) were associated with a poor OS (Table 1). When including radiomic features measured outside the lesions, the model M2 (C-index = 0.77) consisted of 7 features: a high dissemination of lesions, high TMTV, high FDG-uptake in muscles (SULmean > 0.5), low FDG-uptake in spinal cord (SULmean < 0.9), low FDG-uptake in spleen (SULmean < 1), low FDG-uptake in thyroid (SULmean < 0.7) and low volume of muscles (< 7,239 mL) were associated with a poor OS. Model M2 led to a better stratification of patients into 3 risk groups compared to model M1, by identifying both a group of patients with a very good prognosis (57 low-risk patients; 1y-OS rate: 95%, 2y: 84%, 3y: 78%, Figure 1) and a group with a very poor prognosis (24 high-risk patients; 1y: 63%, 2y: 8%, 3y: 0%). Forty-one patients were classified as intermediate-risk patients (1y: 90%, 2y: 49%, 3y: 25%).
Conclusions: Our findings suggest that the addition of radiomic features measured outside the lesion, in combination with clinical data, TMTV and Dmax, leads to a better stratification of TNBC patients according to OS, and helps identify patients with a very good or very poor prognosis. Further investigations are needed to understand the relationship between these features and patient survival, and to validate these results in an independent cohort.