Lung Disease and Cancer Research
Donation protected
Hello Everyone,
My name is Rodney LaLonde and I'm a Ph.D. student working very hard to try to save millions of lives with my medical research involving using artificial intelligence for early diagnosis of lung disease and lung cancer. Unfortunately computer science requires computer hardware to perform this research which can be very expensive. Right now I am making great strides in my goal of early cancer detection and lung disease diagnosis but I really need help getting computer parts to better perform my research. If you can help in any way, you will be a part of research that may save millions of people's lives in the future.
Here is a summary of my research goals and their impacts...
Imaging Biomarker Discovery with Deep Learning for Lung Diseases
Introduction and Motivation:
In 2015, chronic obstructive pulmonary disease (COPD) was the third leading cause of death in the U.S. and globally, estimated to have caused the death of over three million people in the world.[1,2] Lung cancer claimed the second highest number of lives in the U.S. and the fourth most globally with nearly two million deaths.[1,2] Studies are consistently predicting the number of both COPD and lung cancer cases to increase dramatically, with as many as ten million deaths per year predicted from lung cancer alone by 2030.[3,4]
Interstitial lung disease (ILD), while not as globally devastating with 595,000 cases resulting in 471,000 deaths in 2013,5 is among the most technically challenging diagnostically of lung disease cases. There are over 300 different conditions that fall within the ILD group of disorders and only about one in three have a known cause.[6] The diagnostic difficulty in ILD cases stems from current methods requiring the combination of clinical information, radiological (imaging) information, and pathological (biopsy) information to make a successful diagnosis.
While the difficulty to diagnose ILD should not be understated, only 16% of lung cancer cases are diagnosed at an early stage[7] and more than 50% of adults presenting with low pulmonary function were not aware they had COPD.[8] When lung cancer is still localized, the five-year survival rate is 55%; however, for distant tumors it drops to a mere 4%.[7] For COPD, early detection was found to make a “significant health impact” in these patients.[9] Automatic detection and diagnosis of COPD and ILD, as well as researching potential imaging biomarkers for lung cancer in these patients, has the potential to save millions of lives globally.
While the value of human life is immeasurable, this research would be a good financial investment as the National Institutes of Health estimates lung cancer care in 2015 cost the U.S. $13.4 billion and lost productivity due to lung cancer lead to an additional $36.1 billion.[10] Furthermore. the rise of Google DeepMind Health, IBM Watson Health, Enlitic, and others give added confidence to the potential benefits of bringing artificial intelligence to medical imaging.
Shortcomings in Current Methods and My Goal:
Current state-of-the-art methods in lung disease characterization and prediction are insufficient to solve the motivating problems due to their reliance on handcrafted features. These handcrafted features are not discriminative enough to categorize which disorders are present, nor predict cancer development, from imaging information alone. My goal is to develop an automatic diagnosis system using novel deep learning methods and meta-learning to create subject specific models (precision medicine) for the detection/diagnosis of lung diseases and explore their potential relationship with lung cancer using statistical relational learning. I expect this contribution will cause a paradigm shift in lung disease diagnosis, save millions of lives worldwide, and save the U.S. billions of dollars.
Aim 1: Detect and categorize ILD. Hypothesis: Deep learned features will be discriminative enough to categorize ILD disorders from low-dose CT images alone, removing the need for clinical and pathological information. Steps: 1) Create a novel deep learning algorithm for automatic lung parenchyma and lobe segmentation based on [11]. 2) Generate automatic measurements of the airway tree, and lymph nodes. 3) Design a three-dimensional (3D) convolutional neural network (CNN) for general pathology detection using steps 1 and 2.
Aim 2: Detect lung cancer in COPD patients. Hypothesis: A deep CNN, extended from aim 1, will improve detection of lung nodules from low-dose CTs, reduce human efforts in correcting false positives, and help radiologists avoid missing lung nodules during screening. Steps: 1) Design a 3D CNN for lung nodule detection in COPD patients utilizing multi-task learning and a reinforcement-based learning optimization algorithm for optimal neural network design. 2) Design a generative adversarial network to improve the feature representation of detected nodules, thus allowing nodules from cancer and COPD patients to be stratified based on these features. 3) Evaluate the efficacy of the proposed system on a large-scale retrospective COPD data set with known ground truths and a feasibility data set from 500 patients.
Aim 3: Explore relationship with lung cancer. Hypothesis: Studies have shown that COPD is the most important risk factor in lung cancer;4 however, their exact relationship is unknown. Similarly, there has been no unique, quantitative link found so far with ILD. Through the use of statistical relational learning, we hope to discover an imaging biomarker for lung cancer in patients with COPD and/or ILD. Steps: 1) Collect 1000 screening low-dose chest CTs, evaluate them for lung nodules, COPD, and ILD, and apply the proposed algorithms in Aims 1 and 2. 2) Develop a conditional Gaussian process which will classify the extracted (machine learned) features into two groups, while conditioning them to eliminate features which are not uniquely representative of lung cancer in patients with COPD and/or ILD. A Gaussian process is chosen for this pattern discovery and extrapolation because, with the use of a simple and closed-form kernel, it will be possible to perform classification and regression. 3) Test the developed algorithm in step 2 on the data collected at step 1 to see if an imaging biomarker can be identified via this deep-learned Gaussian process.
Intellectual Merit: My three deep learning methods provide several key advancements in medical image analysis, computer vision, and machine learning. 1) Creating a novel deep-learning process to gather measurements of airway trees and lymph nodes can provide a general framework for generating analytics from images. 2) The use of reinforcement learning within a meta-learning framework to design neural network structures will provide key insights into optimal network design across all fields using deep learning. 3) Aggregation of a new generative adversarial network, that will use risk factors to extract diagnostic features, with a Gaussian process could provide a general framework for biomarker discovery in imaging diagnosis.
Broader Impacts: The broader impacts of this proposal are considerable. The most significant impact is the potential improved well-being of individuals in society. This proposal has the potential to not only save millions of lives and billions of dollars but also to improve the well-being of those living with these diseases. Early diagnosis has shown to dramatically improve these individual’s quality of life and their response to treatments. [9]
References: [1] National Center for Health Statistics. Health, United States, 2016: With Chartbook on Long-term Trends in Health. [2] “The Top 10 Causes of Death.” World Health Organization, World Health Organization, Jan. 2017. [3] “Chronic Obstructive Pulmonary Disease (COPD).” World Health Organization, World Health Organization, Nov. 2016. [4] Durham, A.L., and I.M. Adcock. “The Relationship between COPD and Lung Cancer.” Lung Cancer. 2015. PMC. [5] GBD 2013 Mortality and Causes of Death Collaborators. “Global, Regional, and National Age-Sex Specific All-Cause and Cause-Specific Mortality for 240 Causes of Death, 1990-2013: A Systematic Analysis for the Global Burden of Disease Study 2013.” Lancet. 2015. PMC. [6] “The European Lung White Book: Respiratory Health and Disease in Europe.” Interstitial Lung Diseases, European Respiratory Society. [7] U.S. National Institutes of Health. National Cancer Institute. SEER Cancer Statistics Review, 1975-2013. [8] Obstructive lung disease and low lung function in adults in the United States: data from the National Health and Nutrition Examination Survey 1988-1994. Arch Intern Med. 2000. [9] Csikesz et al. “New Developments in the Assessment of COPD: Early Diagnosis Is Key.” International Journal of Chronic Obstructive Pulmonary Disease 9. PMC. [10] U.S. National Institutes of Health. National Cancer Institute. Cancer Trends Progress Report – Financial Burden of Cancer Care. November, 2015. [11] Harrison, Adam P., et al. “Progressive and Multi-Path Holistically Nested Neural Networks for Pathological Lung Segmentation from CT Images.” MICCAI, 2017.
My name is Rodney LaLonde and I'm a Ph.D. student working very hard to try to save millions of lives with my medical research involving using artificial intelligence for early diagnosis of lung disease and lung cancer. Unfortunately computer science requires computer hardware to perform this research which can be very expensive. Right now I am making great strides in my goal of early cancer detection and lung disease diagnosis but I really need help getting computer parts to better perform my research. If you can help in any way, you will be a part of research that may save millions of people's lives in the future.
Here is a summary of my research goals and their impacts...
Imaging Biomarker Discovery with Deep Learning for Lung Diseases
Introduction and Motivation:
In 2015, chronic obstructive pulmonary disease (COPD) was the third leading cause of death in the U.S. and globally, estimated to have caused the death of over three million people in the world.[1,2] Lung cancer claimed the second highest number of lives in the U.S. and the fourth most globally with nearly two million deaths.[1,2] Studies are consistently predicting the number of both COPD and lung cancer cases to increase dramatically, with as many as ten million deaths per year predicted from lung cancer alone by 2030.[3,4]
Interstitial lung disease (ILD), while not as globally devastating with 595,000 cases resulting in 471,000 deaths in 2013,5 is among the most technically challenging diagnostically of lung disease cases. There are over 300 different conditions that fall within the ILD group of disorders and only about one in three have a known cause.[6] The diagnostic difficulty in ILD cases stems from current methods requiring the combination of clinical information, radiological (imaging) information, and pathological (biopsy) information to make a successful diagnosis.
While the difficulty to diagnose ILD should not be understated, only 16% of lung cancer cases are diagnosed at an early stage[7] and more than 50% of adults presenting with low pulmonary function were not aware they had COPD.[8] When lung cancer is still localized, the five-year survival rate is 55%; however, for distant tumors it drops to a mere 4%.[7] For COPD, early detection was found to make a “significant health impact” in these patients.[9] Automatic detection and diagnosis of COPD and ILD, as well as researching potential imaging biomarkers for lung cancer in these patients, has the potential to save millions of lives globally.
While the value of human life is immeasurable, this research would be a good financial investment as the National Institutes of Health estimates lung cancer care in 2015 cost the U.S. $13.4 billion and lost productivity due to lung cancer lead to an additional $36.1 billion.[10] Furthermore. the rise of Google DeepMind Health, IBM Watson Health, Enlitic, and others give added confidence to the potential benefits of bringing artificial intelligence to medical imaging.
Shortcomings in Current Methods and My Goal:
Current state-of-the-art methods in lung disease characterization and prediction are insufficient to solve the motivating problems due to their reliance on handcrafted features. These handcrafted features are not discriminative enough to categorize which disorders are present, nor predict cancer development, from imaging information alone. My goal is to develop an automatic diagnosis system using novel deep learning methods and meta-learning to create subject specific models (precision medicine) for the detection/diagnosis of lung diseases and explore their potential relationship with lung cancer using statistical relational learning. I expect this contribution will cause a paradigm shift in lung disease diagnosis, save millions of lives worldwide, and save the U.S. billions of dollars.
Aim 1: Detect and categorize ILD. Hypothesis: Deep learned features will be discriminative enough to categorize ILD disorders from low-dose CT images alone, removing the need for clinical and pathological information. Steps: 1) Create a novel deep learning algorithm for automatic lung parenchyma and lobe segmentation based on [11]. 2) Generate automatic measurements of the airway tree, and lymph nodes. 3) Design a three-dimensional (3D) convolutional neural network (CNN) for general pathology detection using steps 1 and 2.
Aim 2: Detect lung cancer in COPD patients. Hypothesis: A deep CNN, extended from aim 1, will improve detection of lung nodules from low-dose CTs, reduce human efforts in correcting false positives, and help radiologists avoid missing lung nodules during screening. Steps: 1) Design a 3D CNN for lung nodule detection in COPD patients utilizing multi-task learning and a reinforcement-based learning optimization algorithm for optimal neural network design. 2) Design a generative adversarial network to improve the feature representation of detected nodules, thus allowing nodules from cancer and COPD patients to be stratified based on these features. 3) Evaluate the efficacy of the proposed system on a large-scale retrospective COPD data set with known ground truths and a feasibility data set from 500 patients.
Aim 3: Explore relationship with lung cancer. Hypothesis: Studies have shown that COPD is the most important risk factor in lung cancer;4 however, their exact relationship is unknown. Similarly, there has been no unique, quantitative link found so far with ILD. Through the use of statistical relational learning, we hope to discover an imaging biomarker for lung cancer in patients with COPD and/or ILD. Steps: 1) Collect 1000 screening low-dose chest CTs, evaluate them for lung nodules, COPD, and ILD, and apply the proposed algorithms in Aims 1 and 2. 2) Develop a conditional Gaussian process which will classify the extracted (machine learned) features into two groups, while conditioning them to eliminate features which are not uniquely representative of lung cancer in patients with COPD and/or ILD. A Gaussian process is chosen for this pattern discovery and extrapolation because, with the use of a simple and closed-form kernel, it will be possible to perform classification and regression. 3) Test the developed algorithm in step 2 on the data collected at step 1 to see if an imaging biomarker can be identified via this deep-learned Gaussian process.
Intellectual Merit: My three deep learning methods provide several key advancements in medical image analysis, computer vision, and machine learning. 1) Creating a novel deep-learning process to gather measurements of airway trees and lymph nodes can provide a general framework for generating analytics from images. 2) The use of reinforcement learning within a meta-learning framework to design neural network structures will provide key insights into optimal network design across all fields using deep learning. 3) Aggregation of a new generative adversarial network, that will use risk factors to extract diagnostic features, with a Gaussian process could provide a general framework for biomarker discovery in imaging diagnosis.
Broader Impacts: The broader impacts of this proposal are considerable. The most significant impact is the potential improved well-being of individuals in society. This proposal has the potential to not only save millions of lives and billions of dollars but also to improve the well-being of those living with these diseases. Early diagnosis has shown to dramatically improve these individual’s quality of life and their response to treatments. [9]
References: [1] National Center for Health Statistics. Health, United States, 2016: With Chartbook on Long-term Trends in Health. [2] “The Top 10 Causes of Death.” World Health Organization, World Health Organization, Jan. 2017. [3] “Chronic Obstructive Pulmonary Disease (COPD).” World Health Organization, World Health Organization, Nov. 2016. [4] Durham, A.L., and I.M. Adcock. “The Relationship between COPD and Lung Cancer.” Lung Cancer. 2015. PMC. [5] GBD 2013 Mortality and Causes of Death Collaborators. “Global, Regional, and National Age-Sex Specific All-Cause and Cause-Specific Mortality for 240 Causes of Death, 1990-2013: A Systematic Analysis for the Global Burden of Disease Study 2013.” Lancet. 2015. PMC. [6] “The European Lung White Book: Respiratory Health and Disease in Europe.” Interstitial Lung Diseases, European Respiratory Society. [7] U.S. National Institutes of Health. National Cancer Institute. SEER Cancer Statistics Review, 1975-2013. [8] Obstructive lung disease and low lung function in adults in the United States: data from the National Health and Nutrition Examination Survey 1988-1994. Arch Intern Med. 2000. [9] Csikesz et al. “New Developments in the Assessment of COPD: Early Diagnosis Is Key.” International Journal of Chronic Obstructive Pulmonary Disease 9. PMC. [10] U.S. National Institutes of Health. National Cancer Institute. Cancer Trends Progress Report – Financial Burden of Cancer Care. November, 2015. [11] Harrison, Adam P., et al. “Progressive and Multi-Path Holistically Nested Neural Networks for Pathological Lung Segmentation from CT Images.” MICCAI, 2017.
Organizer
Rodney LaLonde
Organizer
Orlando, FL