Automated Classification of At-home SARS-CoV-2 Lateral Flow Assay Test Results using Image Matching and Transfer Learning: multiple-pipeline study

IntroductionRapid antigen testing for SARS-CoV-2 is an important tool for the timely diagnosis of COVID-19, especially in at-home settings. However, the interpretation of test results can be subjective and prone to error. We describe an automated image analysis pipeline to accurately classify test types and results without human intervention using a dataset of 51,274 rapid antigen test images across three distinct test card brands. MethodsThe proposed method classifies participant-submitted images into four categories: positive for SARS-CoV-2, negative for SARS-CoV-2, invalid/uncertain, and unclassifiable. The model includes four stages: test card classification and region of interest detection using image-matching algorithms, elimination of invalid results using a developed Siamese neural network, and test result classification using transfer learning. ResultsThe model accuracy was very good for test-card classification (100%), region of interest detection (83.5%), and identification of invalid results ranging from 95.6% to 100% for different test types. Performance of the model for test result classification varied by tests; the models sensitivity, specificity, and precision for Abbott BinaxNOW was 0.761, 0.989, and 0.946, BD Veritor At-Home COVID-19 Test was 0.955, 0.993, and 0.877, and for QuickVue(R) At-Home OTC COVID-19 Test was 0.816, 0.988, and 0.930. ConclusionThe proposed method improved the interpretation of rapid antigen tests, particularly in invalid result detection compared to human-read, and offers a great opportunity for standardization of rapid antigen test interpretation and for providing feedback to participants with invalid tests.