Enhanced by the use of smartphones and social media, the amount of digital image content has exploded in a very short time. Not only is image data omnipresent, but we are continuously presented with unrealistic expectations regarding image quality and processing in general. We can upgrade or alter our bodies in social media with only one click (Supermodel? Or dog face?) According to TV detective stories, a crime can be solved by zooming into the mirroring sunglasses of a random pedestrian. Speaking of zoom, high-precision optical techniques in sports, such as Hawk-Eye camera technology, show how emotionally attached we can get to only a few pixels – and how we trust them unconditionally.
Thus, shouldn’t it be easy to do similar kinds of magic with medical images?
Reality check: no. In an AI process, the people who can confirm this fact are the ones labelling the data. Let’s have a look at their master plan.
See the big picture
The first step to succeed in medical image labelling is applying the widely used approach “from general to particular.” Thanks to extensive literature research and the support of partner radiologists, we tackle elementary questions about pathologies’ characteristic features and image artifacts. In reality, MRI studies do not only contain textbook examples. It is crucial to understand basic concepts and patterns in order to deduct particularities manifested in each person’s individual anatomy.
Our work’s centerpiece: the question of belonging
After the first step of understanding fundamental structures and evolving from general to particular, we have already narrowed down quite a lot of possible pixels belonging to the body part in question. What we are left with are a bunch of pixels having definitely more than fifty shades of gray. So where to draw the line?
While using annotation software that does not compute pixel value interpolation, we are able to work with the original pixel values and their differences. We investigate where each pixel belongs by judging if the value difference to the neighbor pixel exceeds a certain threshold. This threshold depends primarily on the surroundings of the potential body part. It is possible that the threshold is not the same in all directions. For example, a threshold difference is expected if a grayish body part is connected to black pixels on the left side and also grayish pixels on the right side. Most important, ensuring consistent labelling criteria amongst our colleagues is the key to maintaining our high standard.
Constant dripping wears the stone
In order to develop a sophisticated deep learning algorithm, our data scientists rely not only on high-precision image labels, but also on an appropriate number of such labels. Thanks to our cooperation with partner hospitals, numerous anonymized MRI studies are available for this purpose. To succeed in medical image labelling, personal commitment is the key to reduce that pile. It is very important to remember the beauty of labelling, as it can become monotonous while demanding constant concentration. After a full work day of “coloring pixels”, we can proudly feel like Paul Klee or a modern pixel art artist.
But what makes our annotators so special that they follow this procedure instead of falling into the trap of unrealistic ideas about the world of images? They are everyday people who possess an important quality: awareness.