Evaluation of Computer-Aided Lung Nodule Detection Software in Thoracic CT for Riverain Technologies LLC

A reader study with at least ten (10) participating radiologists will be conducted. A localized receiver operating characteristic curve will be used to evaluate radiologists' diagnostic performance (in terms of the trade-off between the sensitivity and specificity when the decision criteria changes) in the detection of lung nodules on lung CTs with and without the usage of the ClearRead CT Insight (CRCTI) system. The time needed for the review and interpretation of each case will also be recorded.

An initial (baseline) interpretation will be made by each of the radiologists based on the Lung CT in its original form. At a minimum of one month later, each radiologists will again interpret the same images viewing the pair of CRCTI CT series: Two sets of CT images (standard with CADe marks and processed with vessel suppression will be presented on either one large monitor or two adjacent monitors.

During the baseline reading the radiologist will mark the location of the actionable nodules and assign a score. The radiologist will also indicate the recommended method of follow-up (Contrast CT, PET-CT, CT Follow-up, Biopsy).

During the second reading session (concurrent read), the radiologist will be presented with a standard appearing CT with computer-aided detection (CADe) marks placed and the vessel suppressed same slice with the vessel suppressed view (right image). The second image, vessel suppressed, will be grayed out until the radiologist move the mouse to the second panel. The radiologist will mark locations. These may or may not correspond to the locations of the CAD markers. As before, the radiologist will assign a level of suspicious to each mark and indicate the need, if any, of an additional diagnostic action (CT Follow-up, Contrast CT, PET-CT, or Biopsy).

Based on the levels of suspicion for each nodule and the associated likelihood ratings, LROC curves will be constructed for both the baseline and the concurrent reads and the significance of any difference will be calculated. The recommendations for further action (CT Follow-up, Contrast CT, PET-CT, or Biopsy) will be used to calculate sensitivity and specificity, PPV and NPV.

Number and types of cases:

Retrospective lung CT image series from approximately 300 patients will be included in the study. Approximately one hundred (100) of the patients will have pathology confirmed cancers and approximately two hundred (200) of the patients will be CTs associated with normal patients. Also included as nodule images are those where the actionable nodule was not acted on at that time, but was detected and acted on based on a subsequent CT. These are the prior images where the nodule can be identified and its location is the same as on the "current" image confirmed by the radiologist expert panel (using a majority of three as the decision criterion).

The selected sample, randomly selected from a larger pool of CT cases will be enriched in the following way:

1. Lung nodules (cancers in this study) will be tumor size T1a (20 mm or less). The proportion of nodules 20 mm or less may be increased since this is where the investigators expect the major impact of this software to be.
2. Non-Solid (ground glass) nodules will be added to the sample (based on availability) to determine the performance of the system on non-solid nodules. For this group, to have sufficient cases, the investigators may have to include benign (non-malignant) non-solid nodules.
3. In this project, the investigators will perform a Machine Test of the ClearRead CT Insight algorithm followed by a reader performance evaluation study. Riverain will provide a system configured with the operating point set to be used for the reader studies and a configuration for an "open" system to be used for machine testing and FROC generation.

Arm 1: a baseline read (no secondary content) and Arm 2: concurrent, CAD augmented read.

1. st Arm: Do baseline (measure time, readers score regions according to action and suspiciousness) - mark all locations of concern
2. nd Arm: Concurrent read (measure time, readers score regions according to action and suspiciousness) - mark all locations of concern

The primary study hypothesis is that the adjunctive use of ClearRead CT Insight is superior to use of standard lung CT images alone, as measured by the area under the LROC curve.

STATISTICAL ANALYSES Accuracy

To evaluate the hypothesis of superiority in terms of improvement in accuracy for ClearRead CT Insight vs. unaided, a mixed effects model (DBM) will be implemented (similar to the model outlined in Dorfman, Berbaum and Metz, 1997), where variance components will be included to account for reader, case, reader by case, reader by modality, case by modality and reader by case by modality. However, it is anticipated that the three-way interaction will be inestimable and will subsequently be dropped from the statistical model. Specifically, the hypothesis to test for superiority of ClearRead CT Insight vs. unaided is:

H0: AUCunaided - AUCClearRead CT Insight ≥ 0.0 vs. HA: AUCunaided - AUCClearRead CT Insight < 0

The AUC of the LROC is the primary endpoint to evaluate accuracy and the test of interest will be a two-sided 95% confidence interval on the effect of modality (i.e. ClearRead CT Insight minus unaided). Significance will be concluded if the upper bound of the two-sided 95% confidence interval does not include zero. If the null hypothesis (H0) is rejected, the alternative hypothesis (HA) is accepted and the superiority of using the ClearRead CT Insight system will be established.

Time

The second co-primary objective is to evaluate reduction in time spent per image for ClearRead CT Insight vs. unaided. Specifically, the hypothesis to test for superiority of ClearRead CT Insight vs. unaided is:

H0= Tunaided - TClearRead CT Insight ≤ 0 vs. HA= Tunaided - TClearRead CT Insight > 0

To evaluate the hypothesis of non-inferiority in terms of improved read time for ClearRead CT Insight vs. unaided, a mixed effects model will be implemented (similar to the model outlined in Dorfman, Berbaum and Metz, 1992), where variance components will be included to account for reader, case, reader by case, reader by modality, case by modality and reader by case by modality. However, it is anticipated that the three-way interaction will be inestimable and will subsequently be dropped from the statistical model.

The read times will be tested using a two-sided 95% confidence interval on the effect of modality (i.e. ClearRead minus unaided). Significance will be concluded if the upper bound of the two-sided 95% confidence interval does not include zero.

With either analysis, the use of the mixed model could be modified to employ a bootstrap sampling approach if the model assumptions of the DBM method have been violated. The upper 95 % confidence limit for the difference in the areas under the curves would be calculated using 10,000 bootstraps in the MultiReader MultiCase (MRMC) ROC method.

POWER AND SAMPLE SIZE

The power to detect differences in the AUC of the LROC curve for the proposed statistical analysis using the current design baseline, i.e. 300 cases, each corresponding image read by 10 readers, was assessed through a simulation study. Specifically, the model outlined above was used to simulate 500 datasets across a range of effect sizes, where power was defined as the proportion of datasets that yielded a significant p-value for testing the fixed effect of modality. The method of Dorfman, Berbaum and Metz, 1992 (DBM) was implemented utilizing statistical mixed model theory with jackknife estimates.

These simulations required assumptions regarding the magnitude of the variance components associated with the different random effects. As a pilot study to obtain estimates of variance components was not conducted, variance component estimates from Riverain study SoftView 510(k) (Record # BSSI-PR-09-00006) that uses a similar technology as the proposed ClearRead CT Insight device were used in all power simulations. Please note that the variance components were scaled by the total variance to represent the proportion of total variance explained by each component. All power estimates are dependent upon the appropriateness of the assumed variance components.