Image-based classification of lymphocyte subsets from routine blood smears reveals distinct immune dynamics after CAR-T infusion

Abstract

Abstract Background Chimeric antigen receptor T-cell (CAR-T) therapy is an established treatment for hematological malignancies, yet tools for immune monitoring remain limited. Current assays such as quantitative PCR and flow cytometry are costly and impractical for serial use. Peripheral blood smears (PBS), routinely obtained in clinical care, represent an underutilized, low-cost resource for immune monitoring. In addition to morphologic assessment, serial PBS enables longitudinal tracking of immune dynamics, potentially offering greater insight into reconstitution and toxicity than isolated timepoints. Aims: To define lymphocyte morphologies emerging after CAR-T infusion and train an AI-based platform to automate their classification. We further aimed to apply this approach to a large, multi-product cohort for longitudinal immune monitoring and correlation with clinical outcomes. Methods: We retrospectively assembled a large, multicenter dataset of patients treated with CD19- and BCMA-directed CAR-T therapies. PBS collected around the time of infusion (d-5 to 30) were digitized via Cellavision, and single lymphocyte images were extracted. A subset of cells was manually annotated by expert hematopathologists into predefined morphologic classes based on nuclear and cytoplasmic features. These annotations were used to train a convolutional neural network, which was then applied to all available cells to generate longitudinal morphologic immune profiles. Results: A training set of 1,500 lymphocytes, derived from PBS collected following CAR-T infusion (days 1–14), was manually annotated by a consensus of hematopathologists based on size, chromatin pattern, cytoplasmic features, and nuclear contour. Pathologists classified cell morphologies into five classes (morphotypes): small round lymphocytes (SRL), granular lymphocytes (GL), large granular lymphocytes (LGL), atypical basophilic lymphocytes (ABL), and disrupted cytoplasm lymphocytes (DCL). A convolutional neural network trained on these annotations reliably classified the morphotypes, achieving strong discriminative performance (AUC = 0.85). The model was then applied to 350,000 cells derived from 11,000 PBS of 543 patients treated with CD19- and BCMA-CAR-Ts (days –5 to +30 relative to infusion). We observed marked heterogeneity in morphotype dynamics over time. Prior to infusion, SRL and GL predominated, whereas ABL, DCL, and LGL emerged only post-infusion. Liso-cel and ide-cel demonstrated sustained near-peak absolute lymphocyte count (ALC) levels through day 30. To explore more granular differences between products, we first identified the median day of peak absolute lymphocyte count (ALC) for each CAR-T product, as peak ALC serves as a surrogate for CAR-T cell expansion. Using this product-specific reference day, we then collected the corresponding values of our five morphotypes for all patients. Pairwise statistical comparisons were performed across products, followed by correction for multiple testing. For example, when comparing Axicel to Lisocel, Lisocel has significantly higher ABLs (p<0.01) and lower DCLs (p<0.05). Comparing Ciltacel to Idecel, Ciltacel has significantly higher SRLs (p<0.05), GLs (p<0.005), LGLs (p<0.001), ABLs (p<0.001), and DCLs (p<0.005). To evaluate clinical utility, we modeled the trajectory of ABL proportion using a longitudinal mixed-effects framework stratified by CRS status. ABL trajectories differed significantly, with CRS patients showing a post-infusion rise peaking at day 5 followed by a sharp decline (p < 0.001). Finally, an XGBoost model integrating clinical features (age, CAR-T product) with image-derived morphologic features predicted complete response in non-Hodgkin lymphoma with a cross-validated AUC of 0.71 (±.02), indicating that morphology-informed profiling is informative for outcome prediction. Conclusions: Peripheral blood smears, a widely available and cost-effective modality, reveal distinct lymphocyte morphologies emerging after CAR-T infusion that can be reproducibly classified using deep learning. These morphologic fingerprints correlate with clinical outcomes and support the further development of PBS-based immune profiling as a scalable approach for real-time monitoring of CAR-T patients.