Semantic features associated with outcome in patients treated with CAR-T therapy

BACKGROUND: Recent development in chimeric antigen receptor (CAR) T-cell therapy has shown tremendous promise in providing a durable response in patients with relapsed/refractory (R/R) Diffuse Large B-cell Lymphoma (DLBCL). It has been shown by prior studies that metabolic tumor volume (MTV) is prognostic of treatment response. However, radiological characteristics observed by a clinical expert in the treatment outcome have not been established. We propose to evaluate the role of radiological semantic traits observed on baseline imaging (PET/CT) in relationship to treatment response to axi-cel treatment for patients with R/R DLBCL. The treatment response is measured as progression-free survival (1-year follow-up). METHODS: We systematically assessed five semantic characteristics (Contour, fusion, 5-PS Deauville, number of nodal sites, side of diaphragm) in baseline PET/CT imaging on patients (n=82) with R/R DLBCL. Briefly, the contour trait assesses irregularity in lesion shape observed on the PET image, taking four levels (round, oval, polygon, other-irregular shapes). The fusion metric scores the presence of multi-focal lesions and takes three levels (absent, present without multiple parts, present with multiple parts). The 5-PS follows the Deauville scoring scheme restricted to three levels (1:PS1, 2: PS2&3, 3: PS4&5). The nodal site will count the number of nodal lesions, taking four levels (1 to 3, ≥4). The side of the diaphragm will evaluate the number of lesions on either side of the diaphragm, taking three levels (1: above, 2: below, 3: both). A clinical radiologist uses the above descriptive scoring scheme and provides an unbiased opinion on a Likert scale for each of the characteristics that were scored independently. We then used hazard ratios computed using a Cox-regression model to assess the relationship between the semantic trait observed at baseline imaging and the axi-cel treatment outcome, measured by the progression-free survival (PFS) with 1-year follow-up. We compared our findings with a baseline model that used metabolic tumor volume (MTV) computed at the patient level. RESULTS: We find that the side of the diaphragm is related to adverse outcome for progression-free survival (hazard ratio over 3.7, p-value =0.01). While other semantics metrics (contour, fusion, PS-5, nodal sites) did not show significance to adverse outcomes. We also find that MTV continues to be significantly related to adverse outcomes (hazard ratio of 2.6, p< 0.01). CONCLUSION: We identified semantic traits observed on radiological scans related to adverse patient outcomes. These radiological semantic metrics provide complementary information to MTV and may serve as a surrogate for treatment response. These semantic trails observed on radiological scans can serve as a prognostic indication to select patients who may respond to advanced therapy.