Detection of western blot protein bands using transfer learning in the assessment of risk of gastric and oesophageal cancer

Abstract

We present a transfer learning technique to encode visual features from an immunoblot assay for the detection of immuno-reactive protein bands to understand the association between H. pylori infection and gastric and oesophageal cancer. A CE-marked immunoblot method (HelicoBlot 2.0) was used to analyse 1500 human serum samples, test strips were scanned and the images segmented to enable a machine-learning algorithm to be developed which could identify protein bands on the strips. A model has been developed to detect protein bands with a performance of 95% AUROC. This novel approach to protein band detection reduces time spend by laboratory experts to interpret results, reduces ambiguity in band classification, and can be applied to large-scale epidemiological studies investigating the relationship between infectious disease and risk of cancer.