Abstract 919P
Background
Lung Cancer is among the most common cancer types, and is the leading cause of cancer deaths worldwide. Lung cancer screening penetration remains low with average inclusion rates in 2020 of 6.5%. Artificial Intelligence (AI) tech-based diagnostics stand to help improving this rate by providing more accurate image analysis and patient management.
Methods
An “end-to-end” model designed to detect and characterize the malignancy risk of suspicious pulmonary lesions found on LDCT lung screening scans. The model uses a LDCT as input and provides the location and risk probability for each lesion detected in the scan. The model was trained (lesions: 388 Cancer, 12.332 Benign) and tested (lesions: 151 Cancer, 4.531 Benign) on a subset of the NLST dataset for which lesions were annotated by a team of radiologists. Given that AI explainability is often cited as a concern to clinicians, an evaluation of the clinical parameters that influenced the predictions of the models was performed.
Results
When using a size threshold of 3mm maximal axial diameter the overall lesion level performance of the combined detection and characterization reached an AUC of 0.976 with 94.7% sensitivity and 93.2% specificity. Lesion detection performance was 90.0% sensitivity with an average of 9.8 false positive detections per scan. Results of the clinical feature importance shows that the top 50 features accounting for the models prediction were composed of 28 deep CNN predictions, 11 nodule shape features (quantifying margins, spiculation, size…) and 11 HU texture radiomics (quantifying attenuation, calcification…).
Conclusions
Here we show the evaluation of an AI/ML tech based computer aided detection and characterization (CADe/CADx) with high performances and low false positives. The features of importance for the models were largely based off of deep CNN predictions and were only partially driven by more classical morphological feature extractions.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Median Technologies.
Funding
Median Technologies.
Disclosure
V. Le: Financial Interests, Personal, Full or part-time Employment: Median Technologies. P. Baudot: Financial Interests, Institutional, Full or part-time Employment, as a Data Science Researcher: Median Technologies. C.M. Voyton: Financial Interests, Institutional, Full or part-time Employment, AI SaMD Development Company: Median Technologies. D. Francis: Financial Interests, Personal, Full or part-time Employment: Median Technologies; Financial Interests, Personal, Stocks/Shares: Median Technologies. E. Munoz: Financial Interests, Personal, Full or part-time Employment: Median Technologies. B. Huet: Financial Interests, Institutional, Full or part-time Employment: Median Technologies; Financial Interests, Personal, Stocks/Shares: Median Technologies. All other authors have declared no conflicts of interest.