Leveraging AI Transfer Learning for Rapid and Accurate Identification and Quantification of Cellular Biomarkers in Microscopy Images

Abstract

Many fields of biology research and biomedical diagnostics utilize microscopy imaging to evaluate a broad range of cell types, subcellular structures, and pathology. The data from this imaging is crucial for evaluating experimental outcomes and inferring relationships between relevant biomarkers. With progress in imaging technology, researchers can now view much greater detail and gather larger datasets. Unfortunately, the ability to identify and quantify these biomarkers remains a bottleneck, often requiring significant hand annotation that is both time consuming and potentially introduces intra and inter‐rater error. While some automated solutions exist, they often still require a great deal of user involvement to correct inaccuracies. In recent years, computer vision emerged as an excellent solution to accurately detect regions of interest in microscopy images. The training of deep machine learning object detection models generally requires 1000s of user annotated images. Although this hand‐annotation is still a large upfront time investment, it produces a highly accurate model that provides much faster and more consistent analysis of frequently studied cell types and biomarkers. Regrettably, this method is less useful for researchers studying a novel cell type or biomarker with few training images available. To combat these limitations, we used a machine learning technique known as transfer learning to develop object detection capabilities for novel biomarkers using previously trained models with far fewer annotated images. We tested several different transfer learning methods with two of our existing cellular biomarker detection models, including wisteria floribunda agglutinin (WFA) and Parvalbumin (PV), to create new customized versions for different research labs. Our results indicated that several transfer learning methods worked well, with one performing consistently better than the others. We then tested this best method on a small number of androgen receptor (AR) images, a biomarker for which we did not have an existing detection model, and reached accuracy scores comparable to our original models for other biomarkers. Overall, by utilizing transfer learning we were able to reduce the number of required annotated images from several 1000s to 10s, with comparable or improved detection accuracy. Our results indicate that this computational method will greatly reduce time and resources needed to train accurate object detection models, thus broadening the reach of smart automated analysis to novel biomarkers and smaller datasets. Ultimately, this technology will allow the analysis bottleneck to catch up to the current hardware advances in the field, and rapidly advance biological research progress.