Encrypted login | home

Program Information

Generating Nodule-Like Object Functions for CAD Performance Evaluation in Lung Cancer Screening: Feasibility Study

no image available
A Narita

A Narita*, M Ohkubo , S Wada , Niigata University, Niigata, Niigata


PO-BPC-Exhibit Hall-5 (Saturday, March 5, 2016)  Room: Exhibit Hall

Purpose: To evaluate the performance of computer-aided detection (CAD) in lung cancer computed tomography (CT) screening, we propose a method of using ‘nodule-like object functions’. Nodule-like object functions are calculated from pulmonary nodules in clinical images by deconvolution analysis based on the spatial resolution of a CT scanner. They are applicable for any screening sites to generate computer-simulated (virtual) nodules by convolution based on the spatial resolution of the site’s scanner, allowing site-specific realistic nodule generation. Such virtual nodules will be useful in evaluating site-specific CAD performance. To demonstrate the feasibility of the methodology, we performed a pilot study using spherical objects in a phantom.

Methods: We used a CT test phantom including spherical objects with diameters of 3, 5, 7, and 10 mm and a contrast between sphere and background density of 674 Hounsfield units (HU). The sphere images obtained by a scanner (scanner-1) were deconvolved with the point spread function (PSF) and slice sensitivity profile (SSP) measured for scanner-1; obtained images were referred to as ‘nodule-like object functions’. They were compared with ideal object functions. Next, by convolving the nodule-like object functions with the PSF/SSP of a different scanner (scanner-2), virtual nodules were generated, then compared with real images obtained by scanner-2. The image differences were quantified by the root mean square error (RMSE).

Results: The nodule-like object functions generated from scanner-1 images agreed well with the ideal object function, suggesting the validity of our deconvolution method. Virtual nodules generated from those functions were identical to the nodules in a real image obtained by scanner-2; the RMSE values for 3-, 5-, 7-, and 10-mm-diameter spheres were approximately 11.4, 13.7, 16.1, and 18.4 HU, respectively.

Conclusion: The proposed method was demonstrated to be feasible, and is a potential method for the quality assurance of CAD.

Contact Email: