After an entire childhood of being admonished that writing in books was NOT ALLOWED, I was pretty excited when a high school English teacher said we were going to be taking notes right in the pages of the novels we would be reading. Annotation, she called it. We were encouraged to underline, highlight, circle, and box significant words and passages in the text, as well as record all our brilliant teenage observations in the margins.
In many contexts, an annotation is a note added by way of explanation or commentary. The word “annotation” comes from Latin (adnotare), which means to observe, record, or write something about, on, or near something else. What a useful word! No wonder it has so many varied applications—in textual scholarship, in mathematical expression, in coding, in molecular biology, computational linguistics, and in digital imaging.
“Look mom, cows!”
Just as we teach young children to identify and classify objects in the world around them, praising them when they get it right, so must a neural net be trained. Maria Heinatz, a data scientist at Balzano Informatiks, explains, “It is a fact that one can only recognize what one has seen and learned before. If one does not know what one is looking for, then recognition will be impossible (RE-cognition). That’s exactly how domain-specific annotation learning works.”
In the medical imaging field, annotation is used to draw attention (sometimes using boxes, circles, or arrows) to regions of interest. In the related digital imaging field, the term annotation describes adding metadata to an image in order to train a computer model to recognize certain features. Typically, a medical image annotator performs one of two types of annotation. The first kind, segmentation, involves classifying single pixels. The second kind is classifying a whole image within a dataset. Images are manipulated and encoded in the standard Digital Imaging and Communications in Medicine (DICOM) format. Another widely used format is NIfTI, which produces a 3D image (as opposed to the single slices format of DICOM). Depending on the reader, this format can be manipulated as well (e.g. in 3D).
While being medically trained is not a requirement for performing medical annotation, paying careful attention to discern what body parts usually look like versus how they might appear in the case of pathologies A, B, or C is critical. Says Heinatz, “I think there are people that have more affinity for images/visuals than others. The ‘visual ones’ get very quickly a sense of precision and detail. One that is lacking this ‘visual’ sense will be slower in the learning process. The better an annotator’s understanding of the characteristics of the common AND rare cases, the better his/her annotation is. That’s why quality increases naturally with experience.”
Medical image annotation is the bread and butter of all machine learning development in the healthcare sector. Precisely annotated images are required to train models with accuracy, and an enormous amount of such data is needed for AI solutions to make assessments and predictions with confidence. Without the dedicated work of the thousands of human beings laboring to perform the time-consuming, repetitive tasks of data annotation, it would be impossible to develop the advanced algorithms that may someday be able to diagnose and predict a wide-ranging variety of pathologies.
As AI solutions increasing permeate the healthcare sector, the need for medical annotators is growing. The data annotation tools market is experiencing profound growth. In Europe alone, the market size was valued at USD 390.1 million in 2019 and is anticipated to reflect a growth rate of approximately 27% between 2020 and 2027. This booming demand has created a niche for new companies whose sole function is to provide data labelers trained to use advanced annotation tools.
Image annotators need to have a good eye for detail and a lot of patience. According to one annotator of MRI scans at Balzano Informatiks, “Data annotation is one of the most important aspects of building up an AI-based software. It is time-consuming and rather monotonous work but needs to be accurately done to be able to develop deep neural diagnostic analysis. Without it, there would be no AI-based software solutions.”
The Future of Medical Annotation
Developments in medical annotation will doubtless be driven by demand, and those who are motivated to adapt quickly and learn to use new tools as they become available will lead the way in this growing field. In many non-medical fields and industries, automated image labeling technology is already widely used in order to counteract mental fatigue experienced by annotators. This technology (also known as pre-annotation) is already highly proficient at recognizing common objects, and it is being adapted for identification of domain-specific anatomic structures and other objects seen in various medical contexts. A lot of progress has been made recently with cancer-related automated recognition and labeling, and the technology is being refined for use in other areas of medicine. However, until the machine accuracy rate is consistently very high, human data annotators will continue to be the often unsung heroes of the medical field.