IDH Mutation Classification in Nonenhancing Gliomas: A Comparison of Habitat and Whole‐Tumor Transfer Learning Strategies

ABSTRACT Background Isocitrate dehydrogenase (IDH) mutation status is an important biomarker for the diagnosis and management of nonenhancing gliomas, underscoring the need for noninvasive preoperative classification. Purpose To compare the value of habitat‐based and whole‐tumor strategies in classifying IDH mutation status in nonenhancing gliomas via transfer learning on structural magnetic resonance imaging and subtraction images. Study Type Retrospective. Population Two‐hundred and eighty‐four patients with nonenhancing gliomas, divided into a training set ( n = 198; 44 ± 12 years; 83 females) and a testing set ( n = 86; 46 ± 11 years; 35 females). Field Strength/Sequence 3T, fluid‐attenuated inversion recovery (FLAIR), fast spin‐echo (FSE) T2‐weighted imaging (T2WI), FSE T1‐weighted imaging (T1WI), contrast‐enhanced FSE T1‐weighted imaging (T1CE). Assessment Based on FLAIR, T2WI, T1WI, T1CE, and subtraction images, two regions of interest input strategies were applied to construct transfer learning models, including whole‐tumor strategy and habitat‐based strategy. Model performance was evaluated using the area under curves (AUC) and accuracy (ACC). Finally, the optimal model was combined with clinical variables to develop integrative models. Statistical Tests Continuous variables were analyzed by Student's t test or Wilcoxon rank‐sum test; categorical variables by χ 2 test or Fisher's exact test. Two‐sided p < 0.05 was statistically significant. Results In the whole‐tumor strategy, the subtraction model demonstrated significantly superior performance, achieving training and testing set AUC/ACC of 0.850/0.813 and 0.890/0.884. The habitat‐based strategy significantly outperformed the whole‐tumor strategy, with the T2WI model demonstrating optimal efficacy (training set, AUC/ACC = 0.898/0.899; testing set, AUC/ACC = 0.870/0.849). The integrative model (habitat‐based T2WI + Age + Location) achieved the highest classification performance, with AUCs of 0.923 and 0.947 in the training and testing sets, respectively. Data Conclusion The habitat‐based strategy outperforms the whole‐tumor approach, with the habitat‐based T2WI model achieving optimal classification performance. Integrating age and tumor location into this model can further boost its classification capability. Level of Evidence 3. Technical Efficacy Stage 2.