High Throughput Image Labeling and Lung/Lobar Segmentation on Chest CT using Deep Learning
- Author(s): Wang, Xiaoyong N/A
- Advisor(s): Bui, Alex
- Brown, Matthew
- et al.
Chest CT is the most common modality in thoracic imaging, especially for diagnosis of diffuse lung disease and lung cancer screening. When mining image data from PACS or clinical trials or processing large volumes of data without curation, the relevant scans must be identified among irrelevant or redundant data. Only images acquired with appropriate technical factors, patient positioning, and physiological conditions may be applicable to a particular image processing or machine learning task. Following identification of appropriate images for processing, accurate lung and lobar segmentation is a pre-requisite for the subsequent quantitative image analysis, e.g. air trapping measurement, emphysema scoring, fibrosis scoring, nodule detection, etc. Fully automated segmentation on a diverse spectrum of pathological lungs is still a challenge in clinical practice. Both the labeling and segmentation steps currently require significant manual intervention by image analysts and are prohibitive for large scale processing of big data. The goal of this dissertation is to fully automate the labeling and segmentation tasks in chest CT with high accuracy.
In this dissertation, an image based high throughput labeling pipeline using deep learning was proposed, it aimed to identify anatomical coverage, scan direction, scan posture, lung coverage completeness, contrast usage and breath-hold types. They were posed as different classification problems and some of them required further segmentation and identification of anatomical landmarks. Images of different view planes were used depending on the specific classification problem. All of our models achieved an accuracy > 99% on their respective test sets across different tasks using a research database from multi-center clinical trials. Based on the comprehensive labels from deep learning models, an optimal image series at each time point for a given patient was selected prior to lung and lobar segmentation.
Two fully convolutional networks were proposed to sequentially achieve accurate lung and lobar segmentation. Firstly, a 2D ResNet-101 based network was used for lung segmentation and 575 chest CT scans from multi-center clinical trials were used with radiologist approved lung segmentation. Secondly, a 3D DenseNet based network is applied to segment the 5 lobes and a total of 1280 different CT scans were used with radiologist approved lobar segmentation as ground truth. The dataset included various pathological lung diseases and stratified sampling was used to form the training and test sets with a ratio of 4:1 to ensure a balanced number and type of abnormality were present. Using 5-fold cross validation a mean Dice coefficient of 0.988 � 0.012 and Average Surface Distance of 0.562 � 0.49 mm were achieved by the proposed 2D CNN on lung segmentation. The 3D DenseNet on lobar segmentation achieved a Dice score of 0.959 � 0.087 and an Average surface distance of 0.873 � 0.61mm.