MRI histogram analysis outperforms habitat modelling for predicting 1p/19q codeletion in gliomas

Adding spatially resolved habitat features to a simpler whole-tumor histogram analysis did not improve the prediction of 1p/19q codeletion in IDH-mutant gliomas, according to a retrospective analysis of 122 patients published in the Egyptian Journal of Radiology and Nuclear Medicine. The study, led by Kazuya Sakoda and Shota Nagano, found that whole-region histogram features achieved moderate discrimination on their own, and that the most consistently selected predictors described the relationship between T2-weighted and FLAIR signal distributions rather than the internal architecture of the tumor.

The 2021 World Health Organization classification distinguishes oligodendroglioma, which is IDH-mutant and 1p/19q-codeleted, from astrocytoma, which is IDH-mutant and 1p/19q-non-codeleted. These molecularly distinct entities can look similar on conventional MRI, so preoperative prediction of codeletion status could complement surgical planning and patient counseling. The authors stress, however, that tissue-based molecular testing remains essential for definitive diagnosis, and the new findings explicitly do not support replacing it.

The research team drew on Version 1 of the University of Texas Southwestern Glioma MRI dataset, a public collection of 625 adults with glioma treated between 2006 and 2023. From this source, 126 patients had IDH-mutant grade 2 or 3 tumors with known 1p/19q status. After excluding two cases whose whole-tumor regions contained fewer than 1,000 voxels and two cases where automated segmentation substantially included non-neoplastic white matter abnormalities, the final whole-tumor cohort comprised 122 patients, with 60 non-codeleted and 62 codeleted tumors. Three additional patients were excluded from a parallel tumor-core-like analysis, leaving 119 patients for that comparison.

All analyses used preoperative precontrast T1-weighted, postcontrast T1-weighted, T2-weighted, and FLAIR images. Before public release, the sequences had been registered to a brain atlas, resampled to 1 mm isotropic resolution, and skull stripped using the FeTS platform. The researchers verified spatial consistency and performed no additional registration, bias field correction, or intensity harmonization. Each sequence was normalized by the patient’s own median normal-appearing white matter signal, defined using a white matter probability threshold of 0.90 after excluding a 10-mm region around the tumor. A continuous T2-FLAIR difference map was calculated as normalized T2 signal minus normalized FLAIR signal, a measure motivated by the well-known T2-FLAIR mismatch sign but not intended to classify that visual pattern.

Three models were compared in both the whole-tumor and tumor-core-like regions: a whole-region histogram model using first-order features such as percentiles, skewness, kurtosis, and entropy; a habitat-derived model; and a combined model. Habitat imaging partitions tumors into subregions with distinct multiparametric signal phenotypes, in this case four clusters generated by K-means clustering of the voxelwise normalized signals. To prevent information leakage, habitat generation was performed within each outer training fold, so held-out patients never contributed to signal standardization or centroid estimation.

Validation used 20 repetitions of nested stratified fivefold cross-validation, with inner cross-validation for feature selection and regularization tuning. All preprocessing, feature selection, model tuning, and habitat estimation were confined to the relevant training data. In the whole-tumor analysis, the histogram, habitat, and combined models achieved areas under the receiver operating characteristic curve of 0.747, 0.681, and 0.734, respectively. In the tumor-core-like analysis, the corresponding values were 0.777, 0.734, and 0.766. The tumor-core-like histogram model therefore posted the highest numerical AUC, although the whole-tumor and tumor-core-like cohorts differed by three patients and were not compared directly.

Paired bootstrap analysis with 10,000 patient-level samples quantified the incremental value of habitat features, and the answer was none. The whole-tumor habitat-derived model actually performed worse than the histogram model, with an AUC difference of −0.066 (95% confidence interval, −0.130 to −0.003). The combined model showed no improvement in either region, with differences of −0.013 in the whole-tumor analysis and −0.012 in the tumor-core-like analysis, both with confidence intervals crossing zero. The authors suggest that much of the relevant information may already have been captured by the whole-region T2-FLAIR and T2-weighted distributions, so recalculating first-order features within four subregions may have partitioned the same information rather than adding an independent signal.

Feature selection stability told a consistent story. In the whole-tumor combined models, T2-FLAIR difference kurtosis, normalized T2-weighted kurtosis, and T2-FLAIR difference skewness were selected in 100%, 94%, and 93% of outer training folds, respectively. In the tumor-core-like analysis, T2-FLAIR difference skewness was selected in every fold, alongside normalized postcontrast T1-weighted 90th percentile in 91% and T2-FLAIR difference kurtosis in 89%. The most frequently selected habitat-derived features, such as H4 proportion and habitat Shannon entropy, reached only 80% and 71% in the whole-tumor analysis. The authors note that variation in T2-FLAIR difference skewness and kurtosis may reflect the extent and heterogeneity of incomplete FLAIR suppression, although this interpretation remains hypothetical because quantitative relaxation mapping and spatially matched histopathology were unavailable.

The study also contributed a direct assessment of habitat reproducibility, and here the two regions diverged sharply. Whole-tumor habitat centroids were more stable than tumor-core-like centroids: permutation-based relabeling was required in only 4% of whole-tumor outer folds versus 23% of tumor-core-like folds, with median centroid distances of 0.061 and 0.367, respectively. The authors attribute the lower tumor-core-like stability to the restricted voxel set, which reduced both the number and diversity of available voxels and made the K-means solution more sensitive to the composition of each training fold. Importantly, the higher AUC of the tumor-core-like histogram model does not indicate more reproducible habitat partitioning, and the MRI-defined habitats were interpreted as signal phenotypes rather than histopathologically validated tissue classes.

The authors are candid about the limitations. This was a retrospective, single-center source cohort analysis without external validation, restricted to patients with IDH-mutant grade 2 or 3 gliomas, known 1p/19q status, complete preoperative MRI, adequate region size, and acceptable automated segmentation, which may have introduced selection bias. The sample of 122 patients was modest relative to the 59 habitat-derived and 80 combined candidate features, despite correlation filtering, ridge regularization, and nested validation. Examinations came from multiple scanners and field strengths, but complete sequence-level acquisition parameters were not consistently available, and no scanner-specific harmonization was performed. Automated segmentations were used without manual refinement, and neither the tumor-core-like region nor the habitats were validated against spatially matched histopathology. The number of habitats was fixed at four rather than optimized.

From a clinical perspective, the observed discrimination is insufficient to replace tissue-based molecular testing, and the internally validated results do not establish clinical utility. The authors suggest that whole-region histogram features might eventually provide complementary information when molecular results are pending or tissue sampling is limited, and that the simpler histogram model may be easier to reproduce across institutions because it does not require stable correspondence between intratumoral subregions. The absence of improvement after adding habitat features also suggests that greater model complexity should not be assumed to bring greater diagnostic value. Future studies, the authors conclude, should evaluate external calibration, clinically relevant decision thresholds, and added value beyond conventional radiological assessment rather than AUC alone, using independently processed external datasets.

Subject of Research: Predicting 1p/19q codeletion in IDH-mutant gliomas using MRI histogram and habitat imaging features

Article Title: Multiparametric magnetic resonance imaging histogram and habitat analysis for predicting 1p/19q codeletion in isocitrate dehydrogenase-mutant gliomas

Article References: Sakoda, K., & Nagano, S. (2026). Multiparametric magnetic resonance imaging histogram and habitat analysis for predicting 1p/19q codeletion in isocitrate dehydrogenase-mutant gliomas. Egyptian Journal of Radiology and Nuclear Medicine, 57(1), Article 233. https://doi.org/10.1186/s43055-026-01893-y

Image Credits: AI Generated

DOI: 10.1186/s43055-026-01893-y

Keywords: glioma, 1p/19q codeletion, IDH-mutant, MRI radiomics, habitat imaging, T2-FLAIR mismatch, histogram analysis, oligodendroglioma, astrocytoma, machine learning, WHO classification, cross-validation