10024-CO-1 A MULTICENTER PERFORMANCE COMPARISON STUDY OF AN IDH STATUS PREDICTION MODEL USING THE VISUAL TRANSFORMER MODEL
Satoshi Takahashi, Masamichi Takahashi, Manabu Kinoshita, Mototaka Miyake, Kazuma Kobayashi, Jun Sese, Koichi Ichimura, Yoshitaka Narita, Ryuji Hamamoto, Consortium Glioma Molecular Diagnosis- Surgery
- Oncology
- Neurology (clinical)
Abstract
PURPOSE
To compare the performance and explainability of the visual transformer (ViT) and convolutional neural network (CNN) architectures in predicting genomic mutations from brain MRI.
METHODS
The performances of the ViT and CNN classification models in predicting the IDH mutation status of gliomas were compared. The two models were fine-tuned on the TCIA dataset. The fine-tuned models were evaluated on the TCIA dataset and an external independent dataset, namely the Japanese Cohort (JC) dataset. To evaluate their explanatory power, the gradient-weighted class activation mapping (Grad-CAM) visualization of the CNNs model and attention map visualization of the ViT model were compared.
RESULTS
The visual transformer model consistently outperforms the convolutional neural network on both the TCIA and JC datasets (p-value = 0.021, p-value < 0.001, statistically different). The attention map of the ViT model accurately highlighted the tumor, and the Grad-CAM of the CNN model sometimes highlighted non-tumor areas.
CONCLUSION
The ViT model was more robust against differences in the image domain. The ViT model's attention map had superior explainability.