Semi-supervised method for image texture classification of pituitary tumors via CycleGAN and optimized feature
Hong Zhu1,2, Qianhao Fang3, Yihe Huang3
1School of Medical Information, Xuzhou Medical University, Xuzhou, China. zhuhong@xzhmu.edu.cn.
BMC Medical Informatics and Decision Making
|September 10, 2020
Summary
This study introduces an automated method to assess pituitary tumor texture and softness from MRI scans, improving diagnostic efficiency and accuracy for better surgical planning and prognosis.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
- Neurosurgery
Background:
- Preoperative determination of pituitary tumor softness is crucial for surgical planning and prognosis.
- Current methods rely on manual intervention, limiting efficiency and accuracy.
Purpose of the Study:
- To develop an automated method for diagnosing pituitary tumor texture and predicting softness levels.
- To overcome challenges of unbalanced and under-sampled MRI data.
Main Methods:
- Utilized CycleGAN for domain conversion to address unbalanced/under-sampled MRI data.
- Employed a DenseNet-ResNet based Autoencoder for optimized feature extraction.
- Applied a Convolutional Recurrent Neural Network (CRNN) for softness level classification.
Main Results:
- Achieved high accuracy (91.78%) in classifying pituitary tumor softness.
- Demonstrated superior efficiency and accuracy compared to existing methods.
Conclusions:
- The proposed semi-supervised method accurately grades pituitary tumor texture.
- This approach enhances diagnostic efficiency, aiding surgical selection and prognosis.
