cai-xvkun / escc_ml

Deep-learning Radiomics for Classification Modelling

Deep-learning Radiomics techniques for classification modelling

Introduction

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.

Methods

Analysis flowchart.

• Radiological features extracted from the deep learning method and handcrafted radiomics method
• Machine learning methods in model construction
• Model evaluation

Requirement

All requirements are given in requirements.txt

Python requirements

• 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

R requirements

• psych 1.8.12
• combat 2.0
• data.table 1.12.6

Performance

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.

Reference

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

Harmonization of multi-site imaging data with ComBat