Evaluating the Utility and Limitations of Machine Learning Tumor Segmentation for Automated Longitudinal RANO
Peter Kamel1, Ahmed Naeem2, Hamza Salim2
1From the Department of Neuroradiology (P.K., A.N., H.S., S.A., A.M., C.B., M.W., K.S.), Division of Diagnostic Imaging, MD Anderson Cancer Center, Houston, TX and Department of Radiology (M.W.), The University of Texas Medical Branch, Galveston, TX. peterkamelmd.correspondence@gmail.com.
AJNR. American Journal of Neuroradiology
|April 25, 2026
Summary
Machine learning models show promise for brain tumor segmentation and treatment response assessment. However, fully automated longitudinal Response Assessment in Neuro-Oncology (RANO) faces limitations with post-surgical changes and subtle morphological variations.
Area of Science:
- Neuro-oncology
- Medical imaging analysis
- Artificial intelligence in medicine
Background:
- Machine learning segmentation demonstrates high performance in volumetric brain tumor assessment.
- Clinical translation of these methods for end-to-end treatment response evaluation remains unclear.
Purpose of the Study:
- To assess the capabilities and limitations of machine learning in fully automated longitudinal Response Assessment in Neuro-Oncology (RANO).
- To evaluate the accuracy of an nnU-Net model in classifying treatment response based on RANO criteria.
Main Methods:
- Trained an nnU-Net model on 4,162 brain tumor MRIs for segmentation.
- Applied the model to a longitudinal dataset of 91 patients for tumor segmentation and RANO classification.
- Compared automated classifications against expert ground truth, analyzing accuracy, sensitivity, and specificity.
Main Results:
- The model achieved 77.7% accuracy in distinguishing progressive disease (PD) from non-PD.
- Sensitivity for PD was high (92.1%), but specificity was lower (60.4%).
- Performance varied significantly across response categories, with partial response (PR) being the least accurate (14.3%). False positives were often due to post-surgical changes.
Conclusions:
- Automated segmentation models show potential for treatment response assessment in neuro-oncology.
- Current limitations include challenges with post-surgical enhancement and subtle longitudinal changes in tumor morphology and location.
- Further refinement is needed for fully automated, reliable RANO classification.
