Luigi Tommaso Luppino

Background Radiomic models from multiparametric MRI can characterize tumor infiltration within the non-enhancing peritumoral region but remain insufficiently compared with diffusion and perfusion. This study assessed concordance between voxelwise radiomic predictions and these physiological modalities and evaluated prognostic value of infiltration metrics in two external cohorts. Methods UPenn-GBM and UCSF-PDGM datasets were analyzed. Voxelwise radiomic classification generated peritumoral infiltration-probability maps from standard MRI. Fractional anisotropy (FA), dynamic susceptibility contrast–derived relative cerebral blood volume (DSC-rCBV; UPenn), and arterial spin labeling–derived relative cerebral blood flow (ASL-CBF; UCSF) were compared between peritumoral regions classified as high- versus low-infiltration. Radiomic metrics included infiltration burden (voxel fraction exceeding predefined probability thresholds) and radial extent (normalized maximum distance from enhancing margin sustaining high infiltration probability) were quantified, and survival assessed using univariable Cox models. Results Across 872 subjects, high-infiltration regions showed significantly lower FA (median difference: UPenn −0.232; UCSF −0.226; both p < 0.001) and higher perfusion (median difference: UPenn DSC-rCBV + 0.282; UCSF ASL-CBF + 0.565; both p < 0.001) compared with low-infiltration regions. Infiltration burden at the 0.50 threshold demonstrated prognostic value (UPenn hazard ratio (HR) 2.758, 95% confidence interval (CI) 1.189–6.396, p = 0.018; UCSF HR 21.277, 95% CI 6.024–71.429, p < 0.001). Radial extent was also associated with survival (UPenn HR 2.371, 95% CI 1.215–4.625, p = 0.011; UCSF HR 4.405, 95% CI 1.695–11.494, p = 0.002). Conclusions Voxelwise radiomic infiltration mapping from standard MRI aligns with diffusion and perfusion abnormalities and yields prognostic value. These metrics highlight the role of structural radiomics for characterizing non-enhancing infiltrative spread in glioblastoma.

Glioblastoma is characterized by diffuse infiltration, making accurate detection of residual disease essential for improving prognostication and guiding treatment. This study evaluates whether the volume of predicted infiltration, generated by a machine learning (ML) model trained on radiomic features from postoperative magnetic resonance imaging (MRI), is an independent prognostic factor. We analyzed a total of 114 glioblastoma patients, 89 from a retrospective multicenter cohort and 25 from a prospective cohort, who underwent gross total resection and had an early postoperative MRI. A previously published voxel-wise ML model estimated tumor infiltration probability in the non-enhancing peritumoral region using conventional MRI sequences. High-risk of recurrence regions (HRoR) were delineated from the probability maps, and their volumes were quantified. Associations with residual FLAIR volume, clinical variables (age, Karnofsky Performance Status), and survival outcomes (overall survival [OS], progression-free survival [PFS]) were evaluated using Cox regression and Kaplan–Meier analysis. In the retrospective cohort, multivariate Cox modeling confirmed that higher HRoR volume was independently associated with shorter OS (HR = 1.51; 95% CI, 1.12–2.05; p = 0.008), with no association found for PFS. A robust cutoff of 1.6 cm³ stratified patients into high- and low-risk groups with significantly different OS (456 vs. 678 days; p = 0.038). This threshold was validated in a prospective cohort (326 vs. 525 days; p = 0.039). ML-derived HRoR mapping provides independent prognostic value and may improve risk stratification after surgery in glioblastoma. These findings support its potential clinical integration for personalized follow-up and treatment.

Chain-of-Thought (CoT) prompting has emerged as a simple yet powerful strategy to elicit structured reasoning in large language models. This study investigates how CoT prompting influences reasoning behavior and task performance of large/vision–language models (LLMs/VLMs) applied to Earth Observation (EO). We compare an EO-specialized model (Falcon) with three general-purpose models (LLaVA, LLaVA-CoT, and o3) across two datasets: RSVQAxBEN, a large open EO benchmark, and a proprietary aerial dataset from Narvik. Experiments contrast baseline and CoT-style prompts to assess both factual accuracy and reasoning quality, complemented by an LLM-as-judge evaluation. Results show that CoT prompting benefits only the large-scale o3 model, while smaller or mid-scale models experience degraded accuracy—confirming that effective reasoning is an emergent property of scale. CoT adds transparency by revealing how models reason, though its outputs can still be partly opaque due to safety or internal constraints. On Narvik, o3 generalizes well to unseen EO data, but CoT prompting does not improve quantitative accuracy. These findings suggest that CoT currently offers greater value for interpretability than for performance. Future work should explore inference-time perception–reasoning strategies—where an EO model like Falcon provides scene-level facts that guide o3’s reasoning—to improve both trustworthiness and accuracy without retraining.

With the state-of-the-art IceNet model, deep learning has contributed to an important aspect of climate research by leveraging a range of climate inputs to provide accurate forecasts of Arctic sea ice concentration (SIC). The deep learning subfield of eXplainable AI (XAI) has gained enormous attention in order to gauge feature importance of neural networks, for instance by leveraging network gradients. In recent work, an XAI study of the IceNet was conducted, using gradient saliency maps to interrogate its feature importance. A majority of XAI studies provide information about feature importance as revealed by the XAI method, but rarely provide thorough analysis of effects from reducing the number of input variables. In this paper, we train versions of the IceNet with drastically reduced numbers of input features according to results of XAI and investigate the effects on the sea ice predictions, on average and with respect to specific events. Our results provide evidence that the model generally performs better when less features are used, but in case of anomalous events, a larger number of features is beneficial. We believe our thorough study of the IceNet in terms of feature importance revealed by XAI may give inspiration for other deep learning-based problem scenarios and application domains.