Optimizing Lung Cancer Early Detection by LDCT via Computer Vision
Principal Investigator
Name
Kingshuk Das
Degrees
M.D.
Institution
AnimanDx Inc
Position Title
Founder
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-602
Initial CDAS Request Approval
Nov 14, 2019
Title
Optimizing Lung Cancer Early Detection by LDCT via Computer Vision
Summary
According to the National Cancer Institute (NCI) Surveillance, Epidemiology, and End Results (SEER) database (https://seer.cancer.gov/), lung cancer is currently the leading cause of cancer death in both men and women in the U.S., causing more deaths than colorectal, breast, prostate, and liver cancer combined (from 2009 to 2013). The 5-year survival of localized (confined to primary site) lung cancer is 57.4%, 30.8% for regional disease (spread to regional lymph nodes), and 5.2% for distant disease (metastatic); but unfortunately, only 16% of cases happen to be localized at diagnosis, whereas 22% are regional, and 57% distant. Thankfully, the National Lung Screening Trial (NLST) has demonstrated dramatic potential to allow earlier stage diagnoses, studying a cohort of current/past smokers, ages 55-74, greater than 30 pack-years smoking history, and less than 15 years since cessation of smoking. Utilizing low-dose computed tomography (LDCT), the NLST showed an overall 93.7% (98.3% for localized disease) sensitivity for detecting lung cancer, and importantly 63% of cases were localized at diagnosis (24.3% regional, 12.8% distant; https://www.cancer.gov/types/lung/research/nlst). Unfortunately, the excellent sensitivity of the NLST came at a cost of a 23.5% false-positive rate, which equated to only a 3.6% positive predictive value. Therefore, the overwhelming majority of positive screening results in such a cohort would be expected to be false-positives, representing patients at risk for unnecessary follow-up procedures. These follow-ups include invasive diagnostic procedures such as endobronchial ultrasound (EBUS)- or computed tomography (CT)-guided biopsy—both of which carry small but significant risk of morbidity—and serial noninvasive monitoring (imaging), which would at least come at the cost of patient and healthcare resources, as well as patient anxiety.
The significant false-positive rate for lung cancer screening by LDCT offers an excellent opportunity for optimization via computer vision. The author has previously designed a computational digital image analysis algorithm to effectively discern cancer- from noncancer-derived blood-borne nanoparticles (exosomes and microvesicles) isolated from patient plasma specimens. The author proposes to employ an analogous algorithm to distinguish cancerous from noncancerous lung nodules in LDCT screening images, optimized to significantly reduce the current false-positive rate, using the NLST dataset to train and optimize the algorithm. This will enhance the effort spearheaded by the NLST for early diagnosis of lung cancer by reducing the burden of false-positive screening results.
The significant false-positive rate for lung cancer screening by LDCT offers an excellent opportunity for optimization via computer vision. The author has previously designed a computational digital image analysis algorithm to effectively discern cancer- from noncancer-derived blood-borne nanoparticles (exosomes and microvesicles) isolated from patient plasma specimens. The author proposes to employ an analogous algorithm to distinguish cancerous from noncancerous lung nodules in LDCT screening images, optimized to significantly reduce the current false-positive rate, using the NLST dataset to train and optimize the algorithm. This will enhance the effort spearheaded by the NLST for early diagnosis of lung cancer by reducing the burden of false-positive screening results.
Aims
• Can the computer vision algorithm classify lung nodules identified by LDCT screening as cancerous vs. noncancerous with a reduced false-positive rate, while maintaining or improving current overall sensitivity?
• Can the algorithm identify early stage (localized) lung cancer nodules with a reduced false-positive rate, while maintaining or improving current sensitivity?
• What noncancerous lesion characteristics are most problematic for accurate classification by the algorithm, and can the algorithm be optimized to minimize these errors?
Collaborators
N/A