Study
NLST
(Learn more about this study)
Project ID
NLST-848
Initial CDAS Request Approval
Nov 3, 2021
Title
Validation of Automated Lung Nodule Detection, Segmentation, and Matching
Summary
We have developed an automated processing pipeline that leverages state-of-the-art machine learning techniques to support both lung cancer screening and lung nodule tracking workflows. The pipeline consists of an end-to-end chain of algorithms for nodule detection, segmentation, and longitudinal matching. All three of these components comprise a software analysis package that aims to make lung nodule finding, characterization, and monitoring a more efficient and consistent process. Deployment of such a comprehensive lung cancer detection and analysis package within the clinical setting requires extensive validation. An important aspect in validation is to rule out any bias that may have been inadvertently introduced based on the way in which the algorithm was trained. This goal can be accomplished by demonstrating consistent performance across both unseen data sets that were part of a larger training cohort, as well as data sets from entirely novel collections that were completely left out of the development process. While development of the algorithms described above has been conducted with various sources of CT scan data, the NLST data set represents a rich, independent, and low-dose cohort that is well-suited for our evaluation and validation purposes.
Aims
1. We will use the epicenter locations provided in the “Spiral CT Abnormalities” data to evaluate the sensitivity and specificity of our automatic lung nodule detection system and compare this performance against a comparable validation set of scans from other cohorts.
2. We will compare the diameter and attenuation information provided in the “Spiral CT Abnormalities” data against the same measures extracted from our nodule segmentation algorithm. Accuracy measures will be compared against those extracted from standard-dose scans to determine the effects of a low-dose protocol on segmentation performance.
3. For subjects with one or more comparison reads, we will evaluate the capability of automated nodule identification and matching across timepoints. In subjects with multiple CT acquisitions, and for all nodules correctly identified by our algorithm as part of aim 1, we will compare the results of our automated nodule matching with the comparison read associations provided in the “Spiral CT Abnormalities” data.
Collaborators
Xia Huang, VIDA
Yan Yang, VIDA
John D. Newell Jr., VIDA
Rachel Eddy, VIDA
