Development of Ultra High Resolution CT for Differential Diagnosis of Lung Nodules found using Low Dose CT
Principal Investigator
Name
Rasika Rajapakshe
Degrees
Ph.D.
Institution
BC Cancer Agency Sindi Ahluwalia Hawkins Centre for the Southern Interior
Position Title
Senior Medical Physicist
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-703
Initial CDAS Request Approval
Aug 18, 2020
Title
Development of Ultra High Resolution CT for Differential Diagnosis of Lung Nodules found using Low Dose CT
Summary
Lung cancer is the deadliest form of cancer, leading to millions of deaths annually, worldwide. A major problem is that lung cancer often presents at an advanced stage, where the disease is essentially incurable. Thus, early detection is crucial for survival. It is recommended that high risk groups (smokers aged 55-80 with a 30-pack year smoking history) undergo screening. Current screening procedures utilize low-dose computed tomography (LDCT), giving a 3D image of the patient's lungs with minimal ionizing radiation. However, LDCT has a high false-positive rate (up to 25-35%). False-positives lead to patients unnecessarily undergoing invasive follow-up procedures and risking further complications. Additionally, false-positive results carry the burden of psychological stresses on the patient, family, and friends. It has been found that the highest risk group for suffering complications during follow-up procedures are smokers aged 60-69, while the highest risk group for false-positive results are individuals aged 65+. Thus, the exact group that screening is designed to benefit is also at the highest risk of complications due to current screening procedures.
The goal of this project is to increase the sensitivity and specificity of lung cancer screening procedures through the development of Ultra High Resolution Computed Tomography (UHRCT). We propose the use of a deep learning model that will classify images of benign and malignant lung tissue to determine an optimal image resolution for lung cancer screening. This optimal resolution can be used to suggest adaptations to current CT scanners, such that diagnostic features may be resolved within possibly cancerous lung nodules, while delivering a limited amount of ionizing radiation to the patient. This project seeks to develop UHRCT imaging techniques that are capable of such resolution.
The deep learning model that will be used in this project is a convolutional neural network (CNN). CNN’s have been very successful in the field of computer vision, demonstrating pathologist-level classification in the area of medical images. The model will initially be trained on high resolution pathology images, ensuring pathologist-level classification. Pathology image resolution will then be degraded slightly, training the model on lower and lower resolution pathology images, until it can no longer maintain a pathologist-level classification. The state of the model just prior to this performance drop will be saved and subject to transfer learning. This means that the model will be re-trained on high-resolution micro-CT images (previously acquired), ensuring pathologist-level classification. Similarly, micro-CT image resolution will be degraded, training the model each time until performance falls below the pathologist-level threshold. In this way, the lowest possible resolution of CT image that can be classified at a pathologist-level will be determined. This will yield the optimal resolution for lung cancer screening as it yields the highest predictability of nodule status while subjecting the patient to limited ionizing radiation.
The development of UHRCT imaging at this optimal resolution may then be developed. UHRCT would reduce the number of false-positive patients subject to invasive follow-up procedures while being time and cost effective for the health care system.
The goal of this project is to increase the sensitivity and specificity of lung cancer screening procedures through the development of Ultra High Resolution Computed Tomography (UHRCT). We propose the use of a deep learning model that will classify images of benign and malignant lung tissue to determine an optimal image resolution for lung cancer screening. This optimal resolution can be used to suggest adaptations to current CT scanners, such that diagnostic features may be resolved within possibly cancerous lung nodules, while delivering a limited amount of ionizing radiation to the patient. This project seeks to develop UHRCT imaging techniques that are capable of such resolution.
The deep learning model that will be used in this project is a convolutional neural network (CNN). CNN’s have been very successful in the field of computer vision, demonstrating pathologist-level classification in the area of medical images. The model will initially be trained on high resolution pathology images, ensuring pathologist-level classification. Pathology image resolution will then be degraded slightly, training the model on lower and lower resolution pathology images, until it can no longer maintain a pathologist-level classification. The state of the model just prior to this performance drop will be saved and subject to transfer learning. This means that the model will be re-trained on high-resolution micro-CT images (previously acquired), ensuring pathologist-level classification. Similarly, micro-CT image resolution will be degraded, training the model each time until performance falls below the pathologist-level threshold. In this way, the lowest possible resolution of CT image that can be classified at a pathologist-level will be determined. This will yield the optimal resolution for lung cancer screening as it yields the highest predictability of nodule status while subjecting the patient to limited ionizing radiation.
The development of UHRCT imaging at this optimal resolution may then be developed. UHRCT would reduce the number of false-positive patients subject to invasive follow-up procedures while being time and cost effective for the health care system.
Aims
•Increase sensitivity and specificity of lung cancer screening procedures (reduce false-positive and false-negatives).
•Develop Ultra High Resolution Computed Tomography imaging techniques to achieve diagnostic resolution of lung nodules non-invasively.
•Develop a deep learning model (convolutional neural network) capable of pathologist-level classification on pathology and micro-CT images.
•Use this model to determine the optimal image resolution for diagnosis while minimizing ionizing dose to the patient.
•Reduce the number of patients that are subject to unnecessary invasive follow-up procedures.
•Reallocate time and resources that are spent on false-positive patients to those in need.
Collaborators
•Dr. Islam Mohamed - BC Cancer Agency - Sindi Ahluwalia Hawkins Centre for the Southern Interior
•Dr. Qiuyan Li - Kelowna General Hospital