Differentiating Tumor and Edema in Brain Magnetic Resonance Images Using a Convolutional Neural Network
Purpose: Glioblastoma is the most common subset of glioma with a high grade of mortality. Early diagnosis may cause better therapeutic interventionsand brain MRI shows a good performance on tumor localization. Since manual tumor localization is time-consuming, an automatic tumor segmentation is usua...
Gespeichert in:
Veröffentlicht in: | Frontiers in biomedical technologies 2018-06, Vol.5 (1-2) |
---|---|
Hauptverfasser: | , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Purpose: Glioblastoma is the most common subset of glioma with a high grade of mortality. Early diagnosis may cause better therapeutic interventionsand brain MRI shows a good performance on tumor localization. Since manual tumor localization is time-consuming, an automatic tumor segmentation is usually recommended. Convolutional Neural Network (CNN) has a wide range application for machine vision and visual recognition. Materials and Methods: In this study, an automatic brain tumor segmentation based on a fully CNN is presented. This method has been used to localize and differentiate active tumors including high grade and low-grade from edema in multi-modal MRI containing T1 weighted, T1 enhanced, T2 weighted and FLAIR. For assessing the segmentation performance, a dataset was used and divided into train and test subset. Each image was investigated by sliding the window with different sizes contained 5, 10, 15, 20 and 25 pixels. Results: The results showed that increasing the window size improves the segmentation performance in training phase. It had no significant effect on the segmentation performance in testing phase, therefore increasing the window size improved the learnig of the neural network. The training accuracy for the window with 5 pixels size was 81.6% and for the window with 25 pixels was 96.5%. The test accuracy for the window with 5 pixels size was 80.5% and for the window with 25 pixels was 82.8%. Overall, the best segmentation performance of traning dataset was 97.6% and the best test segmentation performance was 89.7%. Conclusion: The result with training dataset shows that increasing the sliding windows size may cause the increment of accuracy, but this increment may not necessarily increase the accuracy of test dataset. |
---|---|
ISSN: | 2345-5837 |