In Reply: Decoding Glioblastoma Heterogeneity: Neuroimaging Meets Machine Learning

To the Editor: We would like to express our sincere appreciation to Rahman and Noor1 for their insightful commentary on our recently published review, “Decoding Glioblastoma Heterogeneity: Neuroimaging Meets Machine Learning.”2 We are pleased to see that our work has stimulated discussion and acknowledge the important points raised regarding further advancements in glioblastoma imaging and machine learning applications. DIFFERENTIATION BETWEEN TRUE PROGRESSION AND PSEUDOPROGRESSION We appreciate the insightful comments regarding the differentiation between true progression and pseudoprogression. However, we would like to clarify that the primary focus of our review was on using advanced neuroimaging and machine learning models for treatment planning in glioblastoma. That being said, we did touch on relevant advancements in this area, particularly the use of convolutional neural networks (CNNs) and long short-term memory models for distinguishing true progression from pseudoprogression using multiparametric MRI. We also highlighted how hybrid machine learning approaches, incorporating spatial and temporal MRI features, can improve diagnostic accuracy. Nonetheless, we recognize that fully integrating such approaches into routine clinical practice remains a challenge because of variability in imaging protocols and the need for further validation in larger cohorts. Future research should focus on longitudinal monitoring of glioblastoma progression using imaging biomarkers, allowing real-time assessment of tumor dynamics and therapy response. This approach could significantly enhance early detection of treatment resistance and refine therapeutic strategies, ultimately improving clinical decision making and patient outcomes. NEED FOR FULLY AUTOMATED CLINICAL WORKFLOWS We recognize the importance of automated workflows for integrating radiomics and machine learning into clinical decision making. Although our review discusses the role of CNNs in glioblastoma classification, we did not explicitly focus on automation strategies. As the commentary rightly points out, achieving fully automated, scalable clinical workflows requires standardization of imaging protocols and artificial intelligence (AI)–based reporting systems that integrate seamlessly into picture archiving and communication systems. Future research should prioritize the development of automated pipeline solutions that ensure reproducibility and scalability in glioblastoma imaging. A key avenue to explore is federated learning, which enables collaborative model training across multiple institutions without data sharing.3 This could help overcome barriers related to data set heterogeneity while ensuring patient privacy and data security. In addition, regulatory approval processes, including those by the US Food and Drug Administration and European Medicines Agency, will be crucial in ensuring widespread clinical adoption of AI-based radiomics tools. Moreover, ensuring AI models generalize across different clinical settings remains a challenge because of variations in imaging acquisition protocols and patient demographics. Addressing these concerns through robust validation studies will be critical. INTEGRATION OF MULTIMODAL IMAGING WITH GENETIC INSIGHTS The commentary highlights the potential of integrating MRI, positron emission tomography (PET), and computed tomography (CT) imaging with genetic and molecular data to enhance glioblastoma stratification. Although our review extensively covers radiogenomics and the role of MRI-derived imaging biomarkers in predicting isocitrate dehydrogenase status, genetic mutations, and survival, we acknowledge that PET and CT modalities were not emphasized. Future efforts should focus on multimodal approaches incorporating PET-based metabolic imaging and CT-based structural analysis to refine noninvasive tumor characterization and enhance precision oncology strategies. Incorporating emerging molecular imaging techniques, such as hyperpolarized 13C MRI, could provide deeper insights into glioblastoma metabolism,4 aiding in the identification of treatment-resistant tumor subpopulations. Although these techniques offer promising avenues for research, it is unlikely that they will extend beyond a research tool within a limited number of specialized groups. Given the complexity and resource-intensive nature of these methods, it is unlikely they will become part of routine patient care. However, their integration into clinical trial frameworks could still pave the way for more personalized glioblastoma treatment in specific research contexts. MACHINE LEARNING–BASED PROGNOSTIC SUBGROUPING The suggestion to include a discussion on machine learning–based prognostic subgrouping aligns with the focus of our review on radiomics-driven survival prediction. We detailed how CNNs and machine learning models can delineate peritumoral invasion zones and predict survival outcomes. However, we appreciate the opportunity to clarify that such models contribute to glioblastoma prognostic subgrouping by stratifying patients based on distinct imaging and molecular signatures.5 We concur that expanding these methodologies to include real-world validation and clinical trial stratification will be essential for future research. The Figure illustrates the proposed AI-driven glioblastoma imaging and prognostic workflow, integrating multimodal neuroimaging, radiomics, and clinical decision support to optimize patient outcomes.FIGURE.: AI-driven glioblastoma imaging and prognostic workflow. Multimodal imaging inputs, including MRI, PET, and CT, provide rich data sets for radiomics feature extraction. Machine learning models analyze these data to segment tumors, predict genetic mutations, and assess survival prognosis. Clinical decision support integrates AI-driven risk stratification and PACS automation to aid in distinguishing true progression from pseudoprogression. The combined insights lead to optimized outcome prediction and personalized treatment strategies, ultimately improving glioblastoma management and patient care. CNN, convolutional neural network; CT, computed tomography; DTI, diffusion tensor imaging; FLAIR, fluid-attenuated inversion-recovery; IDH, isocitrate dehydrogenase; PACS, picture archiving and communication systems; PET, positron emission tomography.In this context, integration of structural and functional imaging markers with deep learning frameworks could further refine patient stratification,6 potentially allowing for personalized therapeutic strategies tailored to individual tumor characteristics. The inclusion of prospective multicenter studies will be key to validating these models and ensuring their applicability across diverse patient populations. CONCLUSION We are grateful for the thoughtful critique provided, which has highlighted key areas for future exploration in glioblastoma imaging research. Although our review addressed many of these aspects, we acknowledge the need for further discussion on automation strategies and multimodal integration. Future efforts should aim at bridging advanced imaging technologies with clinical applications to maximize their impact on patient care. In addition, we encourage multi-institutional collaborations to accelerate the clinical validation of AI-driven glioblastoma imaging workflows. We also emphasize that these advancements in AI-driven neuroimaging hold the potential to reduce invasive procedures, enhance early diagnosis, and guide individualized treatment plans, ultimately improving patient outcomes and quality of life. We welcome continued dialog on these topics and look forward to future advancements in neuroimaging and machine learning applications for glioblastoma management.