Background:
Immunotherapy assessment in glioma is difficult because treatment triggers complex biological events, including apoptosis, necrosis, edema, vascular changes, and immune infiltration, which often occur without measurable tumor size change, rendering standard criteria (RANO/RECIST) inadequate especially in the context of therapies that exhibit pseudoprogression [1]. While multimodal imaging can detect these processes, each modality alone is hard to interpret, given how rapidly and variably parameters shift across tumors and therapies it is essential to deploy analysis frameworks that are sufficiently advanced to capture the biological complexity and offer reliable therapeutic assessment after immune therapies.

 

Hypothesis/Objective:
A longitudinal multimodal study employing MRI and SPECT/CT across two glioma models and immune therapies, was designed and implemented to test a framework (depicted schematically in Panel A) with the objective of demonstrating the potential of a two-stage multiparametric analysis strategy that can discriminate responder/non-responders, treated/non-treated and can be deployed successfully regardless of tumor or therapy type while also providing insight into the complex underlying biology of tumor progression. Stage 1 would provide a broad statistical screen of biologically-relevant imaging parameters allowing selection of biologically relevant imaging parameters that can distinguish treated from control tumors, followed by Stage 2 consisting in a granular spatio-temporal analysis of the most informative readouts using a multivariate composite approach anchored to a comprehensive ex-vivo histopathology panel.

 

Methods:
Two intracranial glioma models were used: GL261 (C57BL/6 mice, n=8) treated with anti-PD-1 and control IgG, and GBM6 (humanized NSG MHCI&II KO mice, n=12) treated with a bi-specific T-cell engager (BTE) [2,3] or saline (control). MRI (7T Bruker BioSpec; multi-echo spin-echo for T2 mapping) and SPECT/CT ([99mTc]-Duramycin, targeting apoptosis) [4] were acquired at days 0 and 3 (GL261) or days 0, 3, and 6 (GBM6). Multimodal co-registration (~100–200 µm accuracy) enabled voxel-wise tumor analysis. Stage 1 screened anatomic, T2 microstructural, and apoptotic parameters to identify treated/control discriminators; Stage 2 examined spatial (rim vs. core, SUV–T2 coupling, intra-tumoral heterogeneity) and temporal features via a multivariate composite approach. Ex-vivo validation included duramycin-Cy5, cleaved-PARP, TUNEL, Ki67, CD3, CD11b, and CD31.

 

Results:
In-vivo Duramycin-SPECT and T2 data exhibited post-treatment intratumor focal/regional increases in both models. Despite detection of regional hot-spots, whole-tumor averages did not translate into reliable therapeutic metrics (see Panel B). Therapy-induced apoptosis was detected ex-vivo in treated but not in control mice using an optically labeled agent (duramycin-Cy5) and stained tissue sections for cleaved PARP and TUNEL (representative images in persuasive section). In-vivo Duramycin and T2 microstructure signals exhibited spatially heterogeneous increases through Day 3 and 7 in both models. Dynamic spatiotemporal features were consistent with the molecular tracer and microstructure integrating the downstream tissue consequences of the early apoptotic event). The proposed two-stage analysis, focusing on T2 microstructure and apoptotic-uptake signals, identified these parameters as reliable discriminators of treated-versus-control, whereas size and/or peak SUV were not. Notably, the shape of the T2 intra-tumor distribution, its dispersion, and high-T2 tail (% voxels >60 ms), not merely the median, was itself a discriminator reflecting an emerging “long-T2 apoptotic/edematous” subpopulation obscured by mean averaging. Granular analysis identified regional signals (rim vs. core) and a new parameter, voxel SUV–T2 coupling, as more robust predictors of therapeutic response. The multivariate composite for each group, separated arms with no overlap in both models unambiguously (GL261: Cohen’s d=+4.36, Mann–Whitney p=0.029; GBM6: d=+3.24 [Hedges’ g=+2.73], p=0.057; Day 7) (Panel C). Table 1 lists the relevant parameters selected during Stage 1 along with their underlying biological significance and Table 2 lists the composite parameters that were identified and validated at Stage 2 in terms of their ability to differentiate treated mice from untreated ones.

 

Conclusion:
These data suggest that multimodal longitudinal imaging, combined with a biology-driven two-stage framework, can capture the spatiotemporal complexity of immunotherapy response across distinct glioma models and treatments, thereby offering interpretable readouts beyond purely data-driven approaches (i.e. Radiomics) [5]. Incorporating additional modalities targeting complementary biological features could further enhance therapeutic assessment and support personalized, clinically translatable strategies.

 

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
Panel A : Representative schematic depicting the proposed analysis strategy. The upper panel lists glioma-relevant biological processes alongside the corresponding imaging techniques targeting them (some techniques provide readouts linked to combinations of biological processes). The corresponding ex vivo validation assays are also listed. The asterisk ‘ * ‘ and the yellow outline identify the measurements reported in this abstract, whereas the other methods shown are expected to improve the efficiency of our therapeutic assessment (ongoing work). The bottom panel describes the dual-stage approach which relies on multiparametric imaging to discriminate responders/non-responders, treated/untreated.

Panel B: Representative set of images from GBM6 treated (top row) and untreated (bottom row) showing 2D anatomical images with superimposed e corresponding tumor T2 and nSUV parametric maps at Day 3 and 6 post-treatment. The observed changes exhibit some hot spots but do not reliably translate into whole-tumor 3D averages reflective of therapy as schematically depicted in the lower graphic.

Panel C: A biology-driven multivariate composite separates immunotherapy-treated from control tumors across two independent glioma models, where size and single parameters fail. Per-mouse Stage-2 composite response score for GL261/anti–PD-1 (syngeneic, immunocompetent; n=4/arm) and GBM6/BTE (humanized PDX, bispecific T-cell engager, Day 7; n=4 treated, 3 control). Each point is one animal; horizontal bars are group means; the shaded band marks the treated–control separation gap. The composite is a transparent, low-dimensional sum of z-scored, within-subject (post vs. pre) changes in pre-specified, biologically-motivated readouts — tumor volume, T2 microstructure, and apoptotic [⁹⁹ᵐTc]-Duramycin SPECT uptake and its voxel-wise coupling to T2 — with each feature’s direction fixed a priori by its expected response biology, not fit to the data. In both models the composite separates treated from control with no overlap (GL261: Cohen’s d=+4.36, Mann–Whitney p=0.029; GBM6: d=+3.24 [Hedges’ g=+2.73], p=0.057), whereas tumor size alone is non-significant in both. No single parameter — tumor volume, peak SUV, or whole-tumor mean uptake — distinguishes the arms; separation emerges only from the multivariate, interdependent combination. This is the conceptual departure from conventional radiomics: rather than mining a high-dimensional feature bank for statistical correlates — powerful, but frequently an opaque “black box” with limited biological interpretability and uncertain generalizability (Wang et al., Insights Imaging 2025) — every term here is mechanistically pre-specified and anchored to ex vivo histopathology, yielding a compact response signature whose discriminative power is matched by a transparent biological rationale that generalizes across two distinct tumor biologies and immunotherapy mechanisms.

Table 1: Lists imaging parameters, their discriminating potential and biological significance identified in stage 1 of our approach for both tumor models.

Table 2: Identifies the composite terms with most relevance for discriminating the treated from untreated and the mechanistic information derived that enhances theraepeutic assessment.

 

Author
Daniele Procissi
Northwestern University Feinberg School of Medicine