Medical image segmentation as a method for the preparation of learning data
Medical image segmentation is a seemingly a simple process. A number of factors contribute to achieving high quality end results, including proper preparation of medical input data for segmentation.
In this post – as announced in the previous article – we will focus on the preparation of input data, i.e. medical image data for medical image processing.
Moreover, we will have a look at the medical image segmentation process itself and our methodology as well as give some tips on how to improve this process.
Algorithm training and verification vs. data quality
Preparing the data of the learning set and the test set for machine learning in medical image segmentation is usually a time-consuming and costly process, but at the same time necessary to obtain the desired results.
Large sets of diverse data are needed to train and verify the performance of an artificial intelligence algorithm.
Both the quality of the data and the quality of the segmentations themselves are important here. It is on this input information that the results obtained depend.
Anonymization in the process of preparing imaging data for medical image segmentation
Imaging data acquired for software development should provide complete anonymity to patients. Not only personal data but also health information is considered as sensitive data to be protected.
The use of patient data while maintaining patient privacy is possible by applying anonymization or pseudo-anonymization techniques. These allow data to be encoded according to a special, unique key that has no connection to a specific person. During the anonymization process all data that allows for patient identification and whose anonymization is imposed by current legal regulations are removed.
In the case of imaging data, image pixel data should also be encoded if it illustrates recognizable visual characteristics of the patient that would allow identification – such as facial features.
Imaging data – secure storageMedical software development - including AI medical software - continues to evolve, contributing to new advancements. Click To Tweet
One of these is federated learning, an approach that allows AI models to be trained directly at clinical sites. This eliminates the risk of sensitive data leaking out, and the resulting models are placed in an appropriate location.
Keep in mind that the acquired data must be properly stored to facilitate verification and database communication and archiving. And also, to safely use them to evaluate the potential of artificial intelligence [https://www.sciencedirect.com/science/article/pii/S1120179721000958].
Differentiation of imaging data in medical image segmentation
The acquisition of high-quality medical imaging data is part of the AI medical software development process. But in addition to quality, its diversity is also crucial.
It is important to provide as diverse data as possible in the development process of the learning set selection. It is worth ensuring that the data comes from devices of different manufacturers, but also from different medical centers.
In our case, the segmentation results were positively influenced by adding segmentations from several different authors. Such measures allow to improve the results of the neural network.
The diversity of medical imaging data helps to improve the reliability of the network. It makes the network less sensitive to image artifacts or anatomical anomalies.
Providing quality data is worthwhile
Preparing image data sets, performing input segmentation, and developing network architecture and parameters are labor-intensive processes.
The driving force here is the prospect that the effort put in and the activities performed will create solutions leading to the acquisition of objective data that can provide important diagnostic information.
AI as one of the major trends in medicine
As mentioned in [https://www.sciencedirect.com/science/article/pii/S1120179721000958], many researchers and developers are focusing on the development of methods, tools, platforms, and standards to facilitate the process of providing high-quality medical imaging data, as well as the activities of developing artificial intelligence methods.
Certainly, AI is one of the key trends in modern medicine, although only a few AI-based products are currently certified for medical applications.
Will this be enough to gain the trust of the medical world? The answer to this question will likely come in the near future.
A 2019 report by the Deloitte Center for Health Solutions predicts that in another 20 years, MedTech company models will have changed dramatically and collaboration between organizations will have to be tightened to meet the increasingly complex healthcare needs of patients.
If you want to know more about medical image segmentation, please CONTACT US
Read more about our MEDICAL IMAGE SEGMENTATION
See the previous post: Medical image segmentation as an advancement in medical imaging