- Title Information
- Title
- 3D CNN Transfer Learning in CT Lung Scans for Pulmonary Nodule Classification
- Type of Resource
- text
- Name:
Personal
- Name Part
- Wang, Wei
- Role
- Role Term:
Text
- creator
- Name:
Personal
- Name Part
- Duan, Fenghai
- Role
- Role Term:
Text
- Advisor
- Name:
Personal
- Name Part
- Chrysanthopoulou, Stavroula
- Role
- Role Term:
Text
- Reader
- Name:
Corporate
- Name Part
- Brown University. Department of Biostatistics
- Role
- Role Term:
Text
- sponsor
- Origin Information
- Copyright Date
- 2019
- Physical Description
- Extent
- iv, 29 p.
- digitalOrigin
- born digital
- Note:
thesis
- Thesis (Sc. M.)--Brown University, 2019
- Genre (aat)
- theses
- Abstract
- It is a great burden of labor for radiologists to examine each CT image and recognize the nodule by its distinctive shape since some nodules are inconspicuous. The accuracy of manual examination on CT images is also erratic and highly influenced by radiologists’ experience and subjective sensation. The examine results may be inconsistent among different radiologists.
There is a large literature on how to appropriately train the conventional Convolutional Neural Networks (CNN) with CT images. However, the applications of three-dimensional (3D) CNN in the medical field are still preliminary. This study aims to comprehensively review the development of the conventional CNN on computer vision tasks, analyze the challenges faced by 3D CNN and propose a complete process of training a 3D transfer learning model for the pulmonary nodule classification task.
This study proposes a thorough procedure to preprocess raw CT scans, generate decent inputs for training and visualize all the steps. This study utilizes a trained 3D neural network as the feature extractor and trains the last part of the detection layer as well as the last part of the classification layer and freezes all other layers. The framework of the 3D CNN model trained in this study is based on the U-net. This transfer learning model receives preprocessed lung CT scans as inputs and output the predicted lung cancer probabilities as the classification results.
The 3D CNN is proved to have the competence to ensure accuracy and speed for the lung nodules detection and classification since computer models can objectively check each image with equal quality. The area under the ROC curve (AUC) of this transfer learning model is about 0.92 on the NLST test set. By setting the threshold at 0.35, this model achieves the highest classification accuracy rate of 87.5% on the NLST test set.
The results prove that 3D CNN transfer learning models have the competence to ensure the accuracy for pulmonary nodule classification tasks. The further development of deep networks, the availability of large-scale datasets and the tremendous algorithmic innovations are easing the burden of labor for doctors and enhancing the survival rate for patients.
- Subject
- Topic
- Deep Learning
- Subject
- Topic
- CT Image
- Subject
- Topic
- CNN
- Subject
- Topic
- Transfer Learning
- Language
- Language Term (ISO639-2B)
- English
- Record Information
- Record Content Source (marcorg)
- RPB
- Record Creation Date
(encoding="iso8601")
- 20190603
- Identifier:
DOI
- 10.26300/0aaj-f273
- Access Condition:
rights statement
(href="http://rightsstatements.org/vocab/InC/1.0/")
- In Copyright
- Access Condition:
restriction on access
- All rights reserved. Collection is open to the Brown community for research.