Improved Glioma Grade Prediction with Mean Image Transformation

Accurate prediction of glioma grade is significant for treatment planning. Recently, several convolutional neural networks (CNN) were proposed that were trained on one or more combinations of T1, T2, T1 contrast enhanced (T1CE), and fluid attenuated inversion recovery (FLAIR) sequences. In this work, we explore the potential of training CNN by combining these four sequences to form a mean image that encapsulates core characteristics of tumor. The results indicate that the CNN trained with mean images has an F1-score of 94%, which is 7% higher than the CNN trained with stacked sequences. In addition, we explored the potential of machine learning models that were trained on features extracted from CNN. The results indicate that the ML models trained on features extracted from mean computed CNN has an F1-score of 98%, which is 8% superior compared to the CNN trained on stacked sequences. In addition, the proposed model is 2 times smaller with 5.7 million parameters compared to traditional method which has 11.4 million parameters. These findings highlight the potential of combining the sequences for enhancing the accuracy of glioma grading.