The application of ZFNet model on the MRI diagnosis of glioma
Background Glioma is a kind of intracranial space⁃occupying lesion, which causes heavy disease burden. For glioma patients, early accurate diagnosis and early treatment could effectively prolong the progression free survival. Suspected glioma patients would be firstly examined by MRI for diagnosis....
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Tianjin Huanhu Hospital
2021-03-01
|
| Series: | Chinese Journal of Contemporary Neurology and Neurosurgery |
| Subjects: | |
| Online Access: | http://www.cjcnn.org/index.php/cjcnn/article/view/2290 |
| Summary: | Background Glioma is a kind of intracranial space⁃occupying lesion, which causes heavy disease burden. For glioma patients, early accurate diagnosis and early treatment could effectively prolong the progression free survival. Suspected glioma patients would be firstly examined by MRI for diagnosis. The MRI images were mainly read by radiologists artificially. The diagnosis by different radiologists might be different, and the heavy workload might cause a decline in reading efficiency. In recent years, it is becoming possible to use deep learning technology for medical image recognition and diagnosis. This study used machine learning algorithms related to artificial neural network (ANN) to assist image practitioner in reading the head MRI images of patients with glioma, which may cause labor⁃saving, and could increase the efficiency of reading images and reduce the different results cause by different personal experiencse. Methods These images were from The Cancer Imaging Archive (TCIA) database. Format of these images was DICOM. And they were from 130 adult glioma cases, a total of 40 036 copies. These images were randomly split as training set (28 025 copies) and test set (12 011 copies). Then in each set, images were split as "tumor image" and "normal image" according to medical experts' annotation. ZFNet model, a type of convolutional neural network, was used to build image recognition and classification model. The reinforcement learning curve was draw to observe the trend of accuracy of training changed with the training steps. Put the test set into the model, the overall classification accuracy of all MRI images, the positive predictive value, sensitivity, specificity and F1 ⁃ measure of the tumor images were calculated. At the same time, AlexNet was also used to build a same model to compare with the ZFNet model by the prediction indexes of the classification ability of MRI images (the prediction ability of diagnosis of glioma). Results The training accuracy of ZFNet model was 99.7% after 38 757 steps and of the AlexNet model was 98.23% after 37 984 steps. After testing, the image prediction accuracy of all MRI images of ZFNet model was 84.42% (10 140/12 011), the positive predictive value of prediction of "tumor image" was 80.77% (4817/5964), the sensitivity was 86.93% (4817/5541), the specificity was 82.27% (5323/6470), and the F1⁃measure was 83.74%. The above indexes of AlexNet model were 80.74% (9698/12 011), 77.68% (4529/5830), 81.74% (4529/5541), 79.89% (5169/6470) and 79.66%, respectively. The classification performances of ZFNet model were satisfied and were superior to AlexNet model in each dimension. Conclusions ZFNet model has a good prediction performance ability in glioma brain MRI image classification, and it is able to provide a good technical support for establishing a glioma image aided diagnosis model.
doi：10.3969/j.issn.1672⁃6731.2021.03.006 |
|---|---|
| ISSN: | 1672-6731 1672-6731 |