[Paper Review] Comparative Analysis of 2D and 3D ResNet Architectures for IDH and MGMT Mutation Detection in Glioma Patients

Danial Elyassirad, Benyamin Gheiji+6 more|arXiv (Cornell University)|Dec 30, 2024

Glioma Diagnosis and TreatmentMedicine3 citations

TL;DR

This study compares 2D and 3D ResNet architectures for non-invasive prediction of IDH and MGMT mutation status in glioma patients using T1, post-contrast T1, and FLAIR MRI sequences. The 2D ResNet50 achieved the highest AUROC of 0.9096 for IDH status, outperforming 3D ResNet34 (0.8999), while 3D models failed for MGMT prediction, with AUROCs below 0.5.

ABSTRACT

Gliomas are the most common cause of mortality among primary brain tumors. Molecular markers, including Isocitrate Dehydrogenase (IDH) and O[6]-methylguanine-DNA methyltransferase (MGMT) influence treatment responses and prognosis. Deep learning (DL) models may provide a non-invasive method for predicting the status of these molecular markers. To achieve non-invasive determination of gene mutations in glioma patients, we compare 2D and 3D ResNet models to predict IDH and MGMT status, using T1, post-contrast T1, and FLAIR MRI sequences. USCF glioma dataset was used, which contains 495 patients with known IDH and 410 patients with known MGMT status. The dataset was divided into training (60%), tuning (20%), and test (20%) subsets at the patient level. The 2D models take axial, coronal, and sagittal tumor slices as three separate models. To ensemble the 2D predictions the three different views were combined using logistic regression. Various ResNet architectures (ResNet10, 18, 34, 50, 101, 152) were trained. For the 3D approach, we incorporated the entire brain tumor volume in the ResNet10, 18, and 34 models. After optimizing each model, the models with the lowest tuning loss were selected for further evaluation on the separate test sets. The best-performing models in IDH prediction were the 2D ResNet50, achieving a test area under the receiver operating characteristic curve (AUROC) of 0.9096, and the 3D ResNet34, which reached a test AUROC of 0.8999. For MGMT status prediction, the 2D ResNet152 achieved a test AUROC of 0.6168; however, all 3D models yielded AUROCs less than 0.5. Overall, the study indicated that both 2D and 3D models showed high predictive value for IDH prediction, with slightly better performance in 2D models.

Motivation & Objective

• To evaluate the performance of 2D and 3D ResNet architectures in predicting IDH and MGMT mutation status in glioma patients using non-invasive MRI.
• To determine whether 3D volumetric modeling offers advantages over 2D axial, coronal, and sagittal slice analysis for molecular marker prediction.
• To identify the optimal ResNet architecture and modality combination for high-accuracy, non-invasive prediction of key glioma biomarkers.
• To assess the generalizability and robustness of deep learning models across different MRI sequences and tumor phenotypes.

Proposed method

• Trained multiple 2D ResNet variants (ResNet10, 18, 34, 50, 101, 152) on axial, coronal, and sagittal slices from T1, post-contrast T1, and FLAIR sequences.
• Ensembled 2D predictions from three orthogonal views using logistic regression to improve robustness and generalization.
• Trained 3D ResNet variants (ResNet10, 18, 34) on full 3D tumor volumes extracted from the same MRI sequences.
• Split the USCF glioma dataset (495 IDH, 410 MGMT) into patient-level training (60%), validation (20%), and test (20%) sets to prevent data leakage.
• Selected models with lowest validation loss for final evaluation on the independent test set.
• Evaluated performance using AUROC as the primary metric for both IDH and MGMT status prediction.

Experimental results

Research questions

• RQ1 Which ResNet architecture—2D or 3D—yields higher predictive performance for IDH mutation status in glioma patients?
• RQ2 How does the inclusion of 3D volumetric context improve or hinder the prediction of MGMT status compared to 2D slice-based models?
• RQ3 Can 2D models trained on multi-planar slices (axial, coronal, sagittal) achieve better generalization than 3D models using full 3D volumes?
• RQ4 What is the optimal ResNet depth for IDH and MGMT status prediction using MRI data?
• RQ5 Do 3D ResNet models exhibit significant performance degradation for MGMT prediction, and if so, why?

Key findings

• The 2D ResNet50 achieved the highest test AUROC of 0.9096 for IDH mutation status prediction, outperforming all other models.
• The 3D ResNet34 achieved an AUROC of 0.8999 for IDH status, indicating strong performance but slightly below the best 2D model.
• All 3D ResNet models for MGMT status prediction yielded AUROCs below 0.5, indicating they performed worse than random chance.
• The 2D ResNet152 achieved an AUROC of 0.6168 for MGMT status, representing the best performance among all models for this biomarker.
• Ensembling predictions from three 2D views via logistic regression improved robustness and contributed to the top-performing 2D model.
• The study demonstrates that 2D models are more effective than 3D models for predicting IDH status, while 3D models fail to generalize for MGMT prediction.