In biomedical imaging, objects of interest are often characterised by a high variability in their appearance on images. Only experts can reliably identify objects and provide annotations that could be used for downstream analytics. However, the time for experts to annotate biomedical image data is expensive.
To address the challenge, we decided to train a computer vision model that could learn from expert annotations in real time and made it accessible to users via an intuitive web-based interface.
7
Average time that it takes to perform one model update.
2
Average number of updates required to train the model.
98
Average accuracy of the model performance.
During the first phase of the project, we validated that a computer vision model could learn from only a few scribbles annotating foreground (i.e. objects of interest) and background (i.e. any other objects present on images). The trained model would then successfully propagate annotations to other objects that look similar to what experts have already annotated.
Our team analysed the input data that we need to collect from users to train and update the model, as well as the outputs that we need to visualise to users. Based on the analysis, we designed a workflow that allows users flawlessly interact with the model, and prepared mockups to validate the workflow with users.
We implemented the workflow in our proprietary web-based application working as an AI assistant, and deployed it publicly to demonstrate its performance on 3 datasets consisting of images from light, electron and fluorescence microscopy.
“When we first saw the first results, we could not believe our eyes! The model required only a few scribbles to almost perfectly annotate an image. Instead of spending 30-60 min on a single image, we could get reasonable results in a few seconds.”
Our team developed from scratch an AI assistant that drastically reduces the human effort associated with annotating biomedical images.
The AI assistant is based on a computer vision model that learns from scratch in real time from expert annotations and propagates annotations to segment objects that look similar to what experts have already annotated. It yields 98% accuracy after only 2-3 model updates, where each model update takes on average 7 seconds.
We implemented the AI assistant in a proprietary web-based application with an intuitive minimalistic user interface. It allows users to effectively interact with the model, i.e. provide input data required to train the computer vision model, visualise results, fix errors in the model predictions and annotate objects that the model fails to segment.
The AI assistant drastically reduces the costs associated with annotation of biomedical image data and enables new opportunities in biomedical image analytics.
7
Average time that it takes to perform one model update.
2
Average number of updates required to train the model.
98
Average accuracy of the model performance.