Are Algorithms Ready to Take Over Tumor Response Assessment? Lessons from Glioblastoma Imaging

Introduction

Glioblastoma (GBM) remains one of the most aggressive and fatal brain tumors, with median survival rarely exceeding 15–18 months. Accurate imaging follow-up is essential for assessing treatment response, guiding multidisciplinary decision-making, and identifying disease progression before clinical deterioration. However, manual interpretation of MRIs using Response Assessment in Neuro-Oncology (RANO) or Brain Tumor Reporting and Data System (BT-RADS) criteria can be subjective, time-consuming, and limited in reproducibility.

Recent advances in artificial intelligence (AI), particularly deep learning-based volumetric segmentation, are revolutionizing this landscape. This blog reviews three timely and high-quality studies that each explore a distinct facet of AI-based longitudinal response assessment in GBM.

1. Understanding Tool Performance and Limitations
Study: Evaluating automated longitudinal tumor measurements for glioblastoma response assessment

Takeaway for learners: Automated segmentation tools can approximate expert tumor assessments, but consistency, lesion detectability, and progression thresholds remain major hurdles for clinical use.

Suter et al. evaluated two AI segmentation tools—BraTumIA and HD-GLIO-AUTO—using a curated GBM dataset (LUMIERE) with 80 patients and over 500 longitudinal MRIs. The study mimicked clinical follow-up using both automated and expert-derived measurements.

Both tools showed good agreement with expert measurements (HD-GLIO: 81.1%, BraTumIA: 79.7%) when identifying tumor progression. However, both failed in consistently counting lesions and struggled with detecting non-measurable lesions. Most importantly, the study emphasizes that AI outputs still require human oversight, especially for critical decisions such as determining time-to-progression (TTP).

2. Scaling AI to Routine Practice: Promise and Pitfalls
Study: Development and Evaluation of Automated Artificial Intelligence–Based Brain Tumor Response Assessment in Patients with Glioblastoma

Takeaway for learners: Large-scale implementation of AI-based response assessment is feasible, but matching human-level clinical interpretation—especially in survival prediction—remains a challenge.

In a retrospective study involving 634 GBM patients and over 3,400 MRIs, Zhang et al. developed and evaluated AI-VTRA, an AI-driven volumetric response algorithm based on BT-RADS criteria. This tool compared volumetric tumor progression with standardized radiologist assessments and evaluated survival stratification.

While the AI achieved moderate agreement with human radiologists (macro-F1: 0.587–0.755), its predictive accuracy for survival was slightly inferior (Cox model p=0.012). This illustrates a crucial teaching point: AI must be clinically meaningful—not just mathematically accurate. The real-world impact of AI tools will depend on how well they complement, not replace, human expertise.

3. The Bigger Picture: What the Evidence Says
Study: Automated longitudinal treatment response assessment of brain tumors: A systematic review

Takeaway for learners: While machine learning offers substantial promise for automated brain tumor response assessment, current studies often fall short in generalizability, methodological rigor, and clinical readiness.

This systematic review by Shi et al. analyzed 20 machine learning (ML) studies on longitudinal treatment response in brain tumors. While many studies demonstrated strong technical performance (e.g., Dice similarity coefficients >0.8), the review found substantial risks of bias in patient selection, lack of external validation, and inconsistent use of reference standards.

Importantly, few studies used prospective designs or tested across multiple institutions. The review reinforces a vital concept for trainees: before adopting new AI tools in practice, rigorous validation and generalizability must be demonstrated.

Educational Insights for ACORE EDU Learners

• AI Tools Can Save Time—but Not Replace Judgment
  Automated measurements can enhance consistency, but radiologists must interpret these outputs contextually, especially when patient symptoms or prior surgeries alter appearance.

• Volumetry ≠ Gold Standard (Yet)
  While volumetric analysis better reflects tumor burden than 2D diameter measurements, standardized volumetric thresholds still need real-world validation before widespread adoption.

• Clinical Implementation Requires More than Accuracy
  A tool that predicts survival or stratifies response must do so reliably across institutions, scanners, and populations. That means both technical and clinical validation.

• Think Critically About AI Outputs
  If AI suggests progression but the clinical picture is stable, take a step back. Use AI as a support system—not an authority.

Conclusion: What’s Next?

As AI systems become more integrated into radiology, especially in neuro-oncology, it is crucial for clinicians, trainees, and educators to understand both their capabilities and their limits. The transition from manual, 2D-based tumor response assessment to automated, volumetric analysis is underway—but not yet complete. These studies highlight that while AI can enhance efficiency and consistency, human expertise remains irreplaceable.

For radiology learners, now is the time to become AI-literate. Understand how these tools work, how they are evaluated, and where they can fail. The future of neuro-oncologic imaging will be collaborative: radiologist + algorithm. Being on the frontlines of this shift means embracing innovation—critically, wisely, and always with the patient in mind.