Abstract 605P
Background
Lung cancer is the leading cause of cancer deaths globally. Chest X-Rays is one of the imaging modalities to detect nodule and then followed up with CT for characterizing its risk. The aim of this study is to estimate the performance of two algorithms of qXR- the LN algorithm which detects nodule and Lung Nodule Malignancy Score (LNMS) which provides score for the risk of cancer.
Methods
In this study, 106 paired CXR-CT taken within one week at multiple sites in the US were included. Three radiologists reviewed images, mapping CXR-detected nodules to CT without access to qXR output. They assessed texture, calcification, spiculation, and binarized LungRADS scores at 4A into high and low risk. Majority vote was used for the presence, characterization, and malignancy risk. For case-level analysis, the highest risk was used. Detection performance includes nodule and case-level sensitivity, case-level specificity, and False Positive Per Image (FPPI). The AUC of LNMS and the sensitivity (SN) and specificity (SP) at the highest Youden’s point are reported. Zero was imputed for the nodule qXR LN did not detect correctly.
Results
After truthing, 44 cases (60 nodules) were identified as having at least one nodule and rest without nodules. The mean number of nodules among those with nodules was 1.4 (range: 1- 5). Nodule-level SN of qXR was 85.24 (74.27 – 92.03). qXR detected 79% of the nodules smaller than 10 mm, 67% sized 10-20 mm, and 89% larger than 20 mm. At case level the SN and SP were 81.82 (68.03 – 90.48) and 82.25 (70.95 – 89.79). Overall FPPI was 0.1/scan. 90% (54) of nodules identified in CXR was classified as solid in CT and qXR detected 85% (46). 25% (15) of nodules had spiculation and qXR detected 80% (12), and 90% were non-calcified and qXR detected 83% (45) of these. The AUC, SN, and SP of LNMS in categorising nodules as high risk or low risk was 0.795 (0.662 - 0.927), 82.60 (62.86 – 93.02), 80.95 (59.99 – 92.33) respectively.
Conclusions
The qXR system demonstrated high performance in identifying lung nodules and categorizing their risk and comparable performance to human readers reported in literature. While not a replacement for CT, LNMS may provide additional information for CT recommendations.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
S. Chattoraj, S. Shivabasappa, A. Singh, R. Chamadia, S. Sathyamurthy, D. Robert, J. Jeeson, N.S. Vanpully, S.M.J. Shaikh, R. Bose K, M. Nasery: Financial Interests, Institutional, Full or part-time Employment: Qure.ai Technologies Private Limited. All other authors have declared no conflicts of interest.