Multi-task glioma segmentation and IDH mutation and 1p19q codeletion classification via a deep learning model on multimodal MRI

Abstract

Objectives

To develop a deep learning model for simultaneous segmentation of glioma lesions and classification of IDH mutation and 1p/19q codeletion status using multimodal MRI.

Methods

We employed a CNN model with Encoder-Decoder architecture for segmentation, followed by fully connected layers for classification. The model was trained and validated using the BraTS 2020 dataset (132 examinations with known molecular status, split 80/20). Four MRI sequences iamges (T1, T1ce, T2, FLAIR) were used for analysis. Segmentation performance was evaluated using mean Dice Score (mDS) and mean Intersection over Union (mIoU). Classification was assessed using accuracy, sensitivity, and specificity.

Results

The model achieved the best segmentation performance with all four modalities (mDS validation ​= ​0.73, mIoU validation ​= ​0.62). Among single modalities, FLAIR performed best (mDS validation ​= ​0.56, mIoU validation ​= ​0.44). For classification, the combined four modalities achieved an overall accuracy of 0.98. However, classification precision for IDH and 1p19q was potentially limited by class imbalance.

Conclusion

Our CNN-based Encoder-Decoder model demonstrates the benefit of multimodal MRI for accurate glioma segmentation and shows promising results for molecular subtype classification. Future work will focus on addressing class imbalance and exploring feature integration to enhance classification performance.