2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Monday, December 14, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Molecular Testing, Omics Technologies, Serologic Tests, Bioinformatics, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Combination Therapy, Therapy Sequence, Biological Therapies, Antibody Therapy, Transplantation, Bispecific Antibody Therapy, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Immunotherapy, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Machine Learning, Biological Processes, Apoptosis, CHIP, Genomics, Immune Mechanism, Diseases, Lymphoid Malignancies, Plasma Cell Disorders, Myeloid Malignancies, Measurable Residual Disease, Research, Fundamental Science, Translational Research, Clinical Research, Health Outcomes Research, Real-World Evidence, Study Population, Human, Adult
PAPER

5826 Deep FDG-PET/CT imaging phenotypes capture prognostic information for early treatment failure after BCMA-directed CAR T-cell therapy in multiple myeloma

Larry Grullon-Polanco, N/A, Omkar Bhabad2, Myles Kim, PhD3, Robert Norberg4, Omar Alexis Castaneda Puglianini, MD5, Hien Liu, MD6, Melissa Alsina, MD7, Taiga Nishihori, MD8, Kenneth H. Shain, MD, PhD9, Ariosto Siqueira Silva, PhD10, Rachid Baz, MD11, Brandon Blue, MD12, Ariel Grajales-Cruz, MD13, Doris K. Hansen, MD14, John L Cleveland, PhD15, Frederick L. Locke, MD16, Justyn Y Nakashima, MD17, Nicholas Figura, MD18, Yoganand Balagurunathan, PhD19, Rolf Petter Tonseth20, Aaron Wong21, Simon Harrison, MBBS, FRCPath, MRCP, PhD, FRACP22, Mark Raymond Dowling, PhD23, Ghulam Rasool, PhD24, Ciara Louise L Freeman, MD, PhD25
1 Moffitt Cancer Center, Tampa, Florida, United States
2 Moffitt Cancer Center, Tampa, Florida, United States
3 Moffitt Cancer Center, Tampa, Florida, United States
4 Moffitt Cancer Center, Tampa, Florida, United States
5 Moffitt Cancer Center, Tampa, Florida, United States
6 Moffitt Cancer Center, Tampa, Florida, United States
7 Moffitt Cancer Center, Tampa, Florida, United States
8 Moffitt Cancer Center, Tampa, Florida, United States
9 Moffitt Cancer Center, Tampa, Florida, United States
10 Moffitt Cancer Center, Tampa, Florida, United States
11 Moffitt Cancer Center, Tampa, Florida, United States
12 Moffitt Cancer Center, Tampa, Florida, United States
13 Moffitt Cancer Center, Tampa, Florida, United States
14 Moffitt Cancer Center, Tampa, Florida, United States
15 Moffitt Cancer Center, Tampa, Florida, United States
16 Moffitt Cancer Center, Tampa, Florida, United States
17 Moffitt Cancer Center, Tampa, Florida, United States
18 Moffitt Cancer Center, Tampa, Florida, United States
19 Moffitt Cancer Center, Tampa, Florida, United States
20 BC Cancer Agency, Vancouver, British Columbia, Canada
21 Peter MacCallum Cancer Centre, Melbourne, Australia
22 Peter MacCallum Cancer Centre, Melbourne, Australia
23 Peter MacCallum Cancer Centre, Melbourne, Australia
24 Moffitt Cancer Center, Tampa, Florida, United States
25 Moffitt Cancer Center, Tampa, Florida, United States

Introduction

Despite unprecedented response rates with BCMA-directed CAR-T therapy, a proportion of patients experience early progression. FDG-PET/CT is routinely obtained before CAR-T therapy, yet existing approaches reduce whole-body disease patterns to summary measures such as metabolic tumor volume (MTV) and total lesional glycolysis (TLG). While prognostic, these  are not routinely available and may not fully capture the spatial and biologic complexity of disease. We hypothesized that deep image embeddings derived from routine FDG-PET/CT could capture prognostic phenotypes not represented by conventional measures and improve prediction of progression free survival (PFS).

Methods

This IRB-approved retrospective study (Pro00087316) included 501 patients with multiple myeloma treated with BCMA-directed CAR-T who had evaluable pre-infusion FDG-PET/CT within 60 days before lymphodepleting chemotherapy (median 12 days; range 0–60); no post-infusion or serial imaging was used. All scans were acquired on a single PET/CT platform. The pipeline operates directly on standard DICOM and does not require vendor-specific inputs. PET volumes were converted into four maximum-intensity projections: coronal, sagittal, and paired 45-degree oblique views. We systematically evaluated 3,888 preprocessing, encoder, and modeling configurations before locking the final pipeline: a frozen DINOv2 encoder with registers, liver-normalized images, class-token pooling, four-view decision fusion, and an L2-penalized Cox model after 25-component principal component reduction. Three models were compared: a clinical Cox model, clinical plus derived MTV/TLG, and clinical plus the PET embedding risk score. Primary evaluation was Harrell concordance over ten repeats of five-fold cross-validation with identical folds; paired differences in C-index are reported with 95% confidence intervals from a patient-level bootstrap (2,000 resamples). Secondary analyses examined standalone imaging discrimination and calibration.

Results

The clinical plus PET-embedding model achieved the highest discrimination (C-index 0.699) versus clinical alone (0.679) and clinical plus MTV/TLG (0.684). Compared with clinical alone, the PET embedding increased concordance by 0.0198 (95% CI −0.001 to +0.034), whereas adding conventionally derived MTV/TLG increased concordance by 0.0048 (95% CI −0.002 to +0.011). Across the configuration sweep, view aggregation was the only factor that appreciably changed performance, whereas encoder, masking, and Cox regularization had little effect, with best configurations clustering near C=0.66. A single coronal projection captured the majority of the standalone PET signal (C=0.649), and the adopted four-view decision fusion provided a marginal gain (C=0.657).

Although the embedding added no further discrimination once clinical variables and MTV/TLG were combined (0.0088; 95% CI −0.011 to +0.023), as a standalone imaging summary it exceeded MTV/TLG by 0.0382 (95% CI +0.011 to +0.089). The embedding score was only moderately correlated with MTV/TLG (r≈0.53), and after removing burden the residual signal remained prognostic (C=0.57; 95% CI 0.53 to 0.61), indicating that a general-purpose encoder not specifically trained on PET captures prognostic information beyond conventional measures of disease burden and metabolic activity. The combined model was well calibrated in the large (predicted 31% vs observed 34% event probability at 12 months), with over-dispersion of predicted risk (calibration slope 0.63) correctable by recalibration.

Conclusions

In this 501-patient cohort, deep phenotypes derived from routine FDG-PET/CT identified patients at increased risk of progression following BCMA-directed CAR-T and outperformed conventional metrics as a standalone imaging biomarker, supporting the potential utility of imaging-based risk stratification before CAR-T infusion. While incremental gains beyond comprehensive clinical and PET burden models were modest, the consistent advantage of learned image representations across a large-scale configuration study suggests that clinically relevant prognostic signal remains underutilized within routine PET imaging, and is complementary to rather than duplicative of conventional burden metrics. That these results were achieved with a general-purpose vision foundation model rather than a PET-specific architecture supports further development of hematologic PET foundation models and prospective external validation.

Larry Grullon-Polanco, N/A: No relevant conflicts of interest to declare.