R2U3D Recurrent Residual 3D U-Net for Lung Segmentation
Lesion Synthesis and Multi-Scale UNet for Robust Lung Segmentation
Thesis status: Presented and Approved on December 10, 2020
Current page - thesis.dhaval.science
It took more than a year to get it. It's not a piece of paper - a degree. Neither tangible nor a number. It's a simple grit of giving more than getting. An idea that may help save lives and inspire others.
Something that can make my family proud. My master's thesis has been published online.
Cited 25 times. Twice by the scientists from the U.S. Food and Drug Administration.
A quick overview
Problems
Lung infections have irregularities, variation in intensities, uncertain patterns, lesion textures, and shapes
Infected lung and lung infection segmentation are critical tasks
COVID-19 CT data-scarcity to train AI models
Proposed solutions / Contributions
A novel lesion synthesis algorithm to overcome data-scarcity
Our deep neural network and training strategies
Robust lung segmentation and good infection segmentation
Applicability for COVID-19 diagnosis and other lung diseases
The motive of this research is to develop an algorithm and train AI model on normal lungs CT scans, and apply it to,
Segment infected lungs.
Segment the lung infection from segmented lungs.
This work can help find the infection and know the severity of the COVID-19 positive CT scan.
It can help to simplify the CT scan to further classify the infected volume as COVID-19 positive or negative.
It can help 2D/3D lungs and lung infection annotation.
It can be helpful for transfer learning.
Artificial Intelligence (AI) is growing exponentially with novel computational architectures and their cognitive capabilities. AI helps solve complex problems in medical imaging. Lung segmentation is essential since it processes the volumetric information of the lungs, removes the unnecessary areas of the scan, and segments the actual area of the lungs in CT scan. Deep learning is also called deep structured learning that is a part of machine learning. The thesis focuses on deep learning applications to segment lungs and further develop a novel algorithm to make it robust. Supervised learning requires data to train a deep neural network. Deep learning eliminates feature engineering by progressively extracting complex high-level features from available training data. Recently, the deep learning model, such as U-Net, outperforms other network architectures for biomedical image segmentation. In this thesis, two different deep neural networks based on U-Net are proposed for the lung and lung lesion segmentation tasks. The proposed models integrate convolution into the sophisticated Multiscale Recurrent Residual Neural Network based on U-Net. It helps learn spatial dependencies in different dimensions and increases the propagation of spatial information. One of the proposed deep neural networks is trained on the publicly available dataset – LUNA16 and achieves state-of-the-art performance on both LUNA16 (testing set) and VESSEL12 dataset for lung segmentation. Both deep neural network (DNN) and availability of diverse annotated data make the given deep learning based solution robust and generalized for practical use. Even if having sophisticated DNN, scarcity of annotated data challenges the expected outcomes. We are further applying our research to help address medical imaging problems in the current pandemic. Classification of lung CT scans into COVID-19 positive and negative, is an essential task that eases the diagnosis, especially in the absence of other diagnostic tests. Further, having computational power, cloud infrastructure, and peripherals speed up medical imaging based diagnostic tools. Robust segmentation of COVID-19 infected lungs requires rich labeled data. Accurate pixel-level annotation tasks to generate such data are time-consuming, and that delays data preparation. We propose a novel algorithm to generate lesion-like artificial patterns, and U-Net based deep neural network for robust lung segmentation will further help segment COVID-19 lung infection. The pattern generation algorithm generates 2D and 3D patterns to create an enormous amount of synthetic data. This algorithm and DNN give accurate lung segmentation results for highly infected lungs and further provides infection segmentation. This research applies to the preprocessing stages of different applications of deep learning, medical imaging, and data annotation. The proposed work helps the deep neural network to generalize on a given domain to accomplish robust segmentation results in the absence of exact data. AI should never compete with humans and replace but help and assist and collaborate instead. The future of these proposed methods not only depends upon deep learning methods but also depends upon the continuous collaboration with radiologists and the feedback loop. It will go hand in hand benefiting and assisting radiologists to spend more time understanding the patients.
Proposed
Framework
Proposed Lesion Synthesis Algorithm (high-level view)
Dilation based Morphology
Infection patterns over lung area - Infection generation and their bounding boxes - Lesion Synthesis at Local Level
Demo (GIF)
Different Morphologies with Random Kernels
Infection patterns over lung area - Infection generation and their bounding boxes - Lesion Synthesis at Local Level
Demo (GIF)
The first row shows the output using the AI model that has used the developed algorithm while training.
Testing dataset: COVID-19 CT scans
Cite
Kadia, Dhaval. "Advanced UNet for 3D Lung Segmentation and Applications." Master's thesis, University of Dayton, 2021. http://rave.ohiolink.edu/etdc/view?acc_num=dayton1619440426233034
Part 1: 3D lung segmentation (99.2% accuracy) is shown in detail in the IEEE paper and poster 1.
thesis-presentation-dhaval-kadia.pdfI imagined achieving the above results in April 2020, and I am grateful to my advisors for supporting my journey to make it real.
My mother attending my thesis defense virtually
My thesis advisors
Dr. Tam Nguyen
and
Dr. Vijayan Asari
Master's Thesis
Dedicated to my respective parents – Dilip Kadia and Hasumati Kadia
Special thanks to my thesis advisors, professors, mentors, family, and friends.
This idea is just the beginning of something I imagined. I hope it might be of any help and will be happy if it is. Feel free to email me at
