Researchers from Skoltech, Philips Research and Goethe University Frankfurt have trained a neural network to identify anomalies in medical images to help physicians go through countless scans in search of pathologies.
The top two rows show images of cars and digits. Given such data, conventional methods are fairly good at spotting anomalies (right) among ordinary cases (left). The bottom two rows show medical scans — these prove to be more difficult. Image Credit: Nina Shvetsova et al./IEEE Access.
The study explains that the new method is adapted to the nature of medical imaging and is highly successful in identifying abnormalities compared to general-purpose solutions.
The study has been reported in the IEEE Access journal.
The detection of image anomalies is a task that forms part of data analysis in several industries. But medical scans pose a specific challenge. It is much simpler for algorithms to find, for example, a car with a flat tire or a broken windshield from a sequence of car pictures compared to identifying which of the X-rays display early signs of pathology in the lungs, similar to the onset of COVID-19 pneumonia.
Medical images are difficult for several reasons. For one thing, the anomalies look very much like the normal case. Cells are cells, and you usually need a trained professional to recognize something’s amiss.
Dmitry Dylov, Study Senior Author and Professor and Head of the Institute’s Computational Imaging Group, Skolkovo Institute of Science and Technology
“Besides that, there’s the shortage of anomaly examples to train neural networks on. Machines are good at something called a two-class problem. That’s when you have two distinct classes, each of them populated with lots of examples for training — like cats and dogs,” continued Dylov.
“With medical scans, the normal case is always grossly overrepresented, with just a few anomalous examples cropping up here and there. And even those tend to be different between themselves, so you just don’t have a well-defined class for abnormalities,” added Dylov.
Dylov’s research group analyzed four datasets of chest X-rays and breast cancer histology microscopy images to verify the versatility of the method throughout various imaging devices. Although the absolute precision and the benefit gained varied extensively and relied heavily on the concerned dataset, the new technique constantly exceeded the traditional solutions in all of the reviewed cases.
What differentiates the new technique from the competitors is that it looks to “perceive” the general perception that a specialist working with the scans may have by determining the exact features that impact the decisions of human annotators.
The study is also very different due to the suggested recipe for standardizing the method to detect the medical image anomaly so that various research groups can compare their models consistently and reproducibly.
We propose to use what’s known as weakly supervised training. Since two clearly defined classes are unavailable, this task usually tends to be treated with unsupervised or out-of-distribution models. That is, the anomalous cases are not identified as such in the training data.
Dmitry Dylov, Study Senior Author and Professor and Head of the Institute’s Computational Imaging Group, Skolkovo Institute of Science and Technology
“However, treating the anomalous class as a complete unknown is actually very strange for a clinical problem, because doctors can always point to a few anomalous examples. So, we showed some abnormal images to the network to unleash the arsenal of weakly supervised methods, and it helped a lot. Even just one anomalous scan for every 200 normal ones goes a long way, and this is quite realistic,” added Dylov.
The researchers note that their method — Deep Perceptual Autoencoders — is simple to carry forward to an extensive range of other medical scans, beyond the two types utilized in the study, since the solution is altered to the general nature of these images. Specifically, it is sensitive to small-scale anomalies and makes use of a few of their instances in training.
We are glad that the Philips-Skoltech partnership enables us to address challenges like this one that are of great relevance to the health care industry. We expect this solution to considerably accelerate the work of histopathologists, radiologists, and other medical professionals facing the tedious task of spotting minute abnormalities in large sets of images.
Irina Fedulova, Study Co-Author and Director of Philips Research Branch, Lomonosov Moscow State University
“By subjecting the scans to preliminary analysis, the obviously unproblematic images can be eliminated, giving the human expert more time to focus on the more ambiguous cases,” added Fedulova.
Journal Reference:
Shvetsova, N., et al. (2021) Anomaly Detection in Medical Imaging With Deep Perceptual Autoencoders. IEEE Access. doi.org/10.1109/ACCESS.2021.3107163.