This is the code for the paper entitled "Computed tomography-based deep-learning prediction of neoadjuvant chemoradiotherapy treatment response in esophageal squamous cell carcinoma"
This code includes feature extraction from pretrained Deep Convolutional Neural Network model and then further model training and validation by machine learning approaches.
Analysis flowchart.
- Radiological features extracted from the deep learning method and handcrafted radiomics method
- Machine learning methods in model construction
- Model evaluation
All requirements are given in requirements.txt
- numpy 1.16.3
- pandas 0.24.2
- scipy 1.2.1
- SimpleITK 1.2.0
- Keras 2.2.4
- scikit-learn 0.21.1
- psych 1.8.12
- combat 2.0
- data.table 1.12.6
The followed table showed the area under the receiver operating characteristic curve (AUC) by
different feature extractorsin the external test cohort
Method | AUC |
---|---|
ResNet50 | 0.805 |
Xception | 0.763 |
VGG16 | 0.648 |
VGG19 | 0.635 |
InceptionV3 | 0.753 |
InceptionResNetV2 | 0.653 |
Radiomics | 0.725 |
The feature maps generated from ResNet50 indicated locations that were important for output
generation (followed figure). Tumoral and peri-tumoral areas of the images were shown to be
valuable for the feature pattern extraction.
Xception: Deep Learning with Depthwise Separable Convolutions
Very Deep Convolutional Networks for Large-Scale Image Recognition
Deep Residual Learning for Image Recognition
Rethinking the Inception Architecture for Computer Vision
Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
Visual Explanations from Deep Networks via Gradient-Based Localization