Development of a Deep Learning Model for Accurate Preoperative Identification of Glioblastoma and Solitary Brain Metastases By Combining Multi-Centre and Multi-Sequence Magnetic Resonance Images
Abstract
Purpose
Development of a deep learning model for accurate preoperative identification of glioblastoma and solitary brain metastases by combining multi-centre and multi-sequence magnetic resonance images and comparison of the performance of different deep learning models.
Methods
Retrospectively collect clinical data and MR images of a total of 221 patients with pathologically confirmed glioblastoma and solitary brain metastases from January 2019 to May 2024 in Provincial Hospital of Shandong First Medical University, of which 160 cases were used as the training set; preprocess the images and label the tumour regions, and input the different MR sequence images individually or in combination to train the deep learning model VisionTransformer (VIT) was used to enhance the data input through ten-fold cross-validation, and the optimal sequence combinations were obtained to train the deep learning models VGG, Resnet, and MobileNetV2; the Loss curves were used to evaluate the degree of fit of the different models, and the accuracy, precision, and recall of the models were evaluated.
Results
In the identification comparison of different MR sequences for discriminatory efficacy, the three sequence combinations of T1-CE, T2, and T2-Flair gained discriminatory efficacy with an accuracy rate of 93.78 and a precision rate of 94.17; after the four deep learning models were inputted into the former sequence combinations, MobileNetV2 had an accuracy rate of 95.63 in internal validation and a precision rate of 95.86, and the external validation had an accuracy rate of 87.18 and precision rate of 87.21, which are the highest among all models.
Conclusion
A combination of multi-sequence MR images and a deep learning model can efficiently identify glioblastoma and solitary brain metastases preoperatively, and the deep learning model MobileNetV2 has the highest efficacy in identifying the two types of tumours.