[Abstract] Artificial Intelligence for Detecting ACVR1 Mutations in Patients with DIPG Using MRI and Clinical Data

Artificial Intelligence for Detecting ACVR1 Mutations in Patients with DIPG Using MRI and Clinical Data

Fahad Khalid1, Thibault Escobar1, Jessica Goya-Outi1, Vincent Frouin2, Nathalie Boddaert3, Jacques Grill4, Frederique Frouin1
 1 LITO U1288, Inserm, Institut Curie – Centre de recherche, Université Paris-Saclay, Orsay, France,
2 GAIA, Neurospin, CEA, Gif-sur-Yvette, France,
3 Hôpital Necker Enfants Malades, AP-HP, IMAGINE Institute, Inserm, Université de Paris, Paris, France,
4 Gustave Roussy, Inserm, Villejuif, France
Presented at ISPNO 2022 in Hambourg


Introduction: ACVR1 mutations are found in about 25% of patients with diffuse intrinsic pontine glioma (DIPG). Recent work has identified the combination of vandetanib and everolimus as a promising therapeutic approach for these patients. We investigate the predictive power of an AI model integrating clinical and radiomic information to predict ACVR1 mutation.

Methods: This retrospective monocentric study includes 65 patients with known ACVR1 status. Patients were scanned at the diagnosis time with at least one of the four structural MRI modalities (pre- and post-contrast T1, T2, FLAIR) and basic clinical information (age and sex) was collected. Radiomic features were extracted within the tumor region from each modality. For each modality, a recursive feature elimination method was used to select the most relevant features. Inside a leave-one-out framework, up to five logistic regression models were built: one per MRI modality and one for the clinical information. The final prediction for each patient was computed as the mean of the probabilities of ACVR1 mutation for the up to 5 different models. Assigning a different weight to clinical data according to age, (more or less than 10 years old) was also tested.

Results: Out of the 65 patients (mean age 7.9±3.7, 15 patients older then 10 years), ACVR1 mutations were identified with a 78% accuracy (sensitivity = 92% and specificity = 75%) in the leave-out-out process. Accounting for the clinical data in the model increase the accuracy to 82% (resp. sensitivity = 86% and specificity = 80%).

Conclusion: The proposed multi model approach compensates for missing MR modalities while taking advantage of all the available information. Our first results suggest that a dedicated model could be developed for younger patients to improve the prediction. The different models will now be tested using additional data coming from the ongoing multicentric BIOMEDE trials.

>See the abstract