Molecular alterations prediction in gliomas via an interpretable deep learning model: a multicentre and retrospective study
Menée à partir de 1 696 images de lames histopathologiques issues du projet "The Cancer Genome Atlas" (877 patients) puis validée à partir de 4 769 images supplémentaires (3 235 patients), cette étude multicentrique évalue la performance d'un modèle d'apprentissage profond interprétable pour prédire la signature moléculaire des gliomes
Background : Molecular profiling of gliomas has a pivotal role in diagnosis, treatment selection, and prognostic assessment. However, it heavily relies on time-consuming and expensive genomic testing, which is largely inaccessible in resource-limited settings. To enable cost-effective and scalable identification of molecular alterations, we developed and validated a foundation model-based interpretable approach to predict key molecular events directly from routine histopathology slides without manual annotation.
Methods : We developed the glioma molecular alterations predictor (GMAP), a foundation model-based approach using 1696 whole-slide images from 877 patients downloaded from the Cancer Genome Atlas. The model was validated on an internal test set (167 whole-slide images from 88 patients) and a grouped external validation set (4602 whole-slide images from 3147 patients; 12 Chinese hospitals and a public dataset, EBRAINS). The performance was primarily evaluated at the patient level by the area under the receiver operating curve (AUROC), accuracy, sensitivity, specificity, and F1 score, with probabilities aggregated across multiple slides per patient by averaging. The interpretability was evaluated through multilevel analysis of high-contribution tiles, and comparative assessment between model-generated heatmaps and corresponding immunohistochemical staining patterns.
Findings : The GMAP reached AUROCs of 0·939 (95% CI 0·865–0·993) for isocitrate dehydrogenase (IDH), 0·955 (0·898–0·992) for the co-deletion of chromosome arms 1p and 19q (1p/19q co-deletion), 0·944 (0·849–1·000) for telomerase reverse transcriptase (TERT), and 0·886 (0·802–0·955) for chromosome 7 gain and chromosome 10 loss (+7/–10) on the internal test set, respectively. In the grouped external validation set, the AUROCs was 0·870 (95% CI 0·857–0·883) for IDH, 0·885 (0·865–0·905) for 1p/19q co-deletion, 0·694 (0·665–0·724) for TERT, and 0·672 (0·615–0·727) for +7/–10. Interpretability analysis showed that GMAP attends to both known and previously unrecognised morphological characteristics associated with molecular alterations.
Interpretation : GMAP offered a technically feasible approach for accurate, fast, and potentially cost-effective identification of molecular alterations in resource-constrained settings. Interpretability analysis revealed model-attended features, which improve the model’s trustworthiness for clinical adoption.
Funding : National Natural Science Foundation of China.
The Lancet Digital Health , article en libre accès, 2026