The AI NerveTrack Model training procedure for Samsung Medison is streamlined and made simpler by Intel Geti Platform.
NerveTrack, a cutting-edge ultrasonography feature from Samsung Medison, allows real-time nerve structure identification while anaesthesia is being administered. Doctors and AI technologists worked together to create the AI models that provide NerveTrack its accuracy and speed.
Thousands of annotated ultrasound reference images are needed to train NerveTrack’s deep learning inference models. Due in part to the fact that annotation tools and methods are usually created for computer engineers and data scientists who will eventually use the images to train the models, the process of annotating images can be challenging and time-consuming for physicians who have years of medical training and experience in recognising the tiny, elusive nerve structures.
Samsung Medison collaborated with the Intel Geti platform to enhance the teams’ collaboration and annotation and modelling workflow. With the help of the computer vision AI platform’s user-friendly interface, a small team of doctors was able to annotate tens of thousands of images in a matter of weeks before sending them to the AI engineer team for model training.
The Difficulty of Including Medical Professionals in the Training of AI Models
Using its high-resolution ultrasound scanners, Samsung Medison, a multinational manufacturer of medical equipment, created NerveTrack as a real-time assistance for physicians doing ultrasound-guided regional anaesthesia (UGRA) and pain management. As they insert their needles at the nerve spot, anaesthesiologists and other medical professionals observe the ultrasound display. NerveTrack enables the physicians to more rapidly and precisely identify the nerve structures.
Real-time nerve structure identification is made possible by the deep learning inference models that underpin the NerveTrack model. NerveTrack’s proprietary AI inference algorithm was initially taught to identify nerves solely in the wrist. Samsung Medison then concentrated on enhancing its capabilities so that an AI inference model could also recognise peripheral nerve structures in the neck, shoulder, and elbow in order to increase its applicability.
Because peripheral nerves are so thin and tiny, they are particularly challenging to find. The nerve itself may be obscured by “noise” or artefacts in an ultrasound picture. Because nerves differ in appearance, it is crucial to rely on the knowledge of neurologists and anaesthesiologists throughout the image annotation process in order to recognise nerves in a wide range of images. Without the annotations, a deep learning-based AI model might not be able to distinguish the small, biological structures from noise in the image or accurately identify the nerves. Because nerves can move in relation to other body structures, ultrasound images are usually recorded as a sequence of video frames.
Samsung Medison required hundreds of thousands of ultrasound pictures to be tagged as references in order to train each NerveTrack model because of the tiny size, complexity, and unpredictability of the nerve structures. However, because these computer specialists do not have a thorough understanding of human anatomy, AI developers are unable to annotate the photos themselves. Instead, medical professionals who have studied and worked with nerve structures for years, such neurosurgeons, anaesthesiologists, and other doctors, must annotate the information.
Finding time with the busy physicians and specialists who have the necessary skills for picture annotation was a major obstacle for Samsung Medison’s NerveTrack application expansion.
The Solution: Intel Geti Platform Puts Power in Doctors Hands
Samsung Medison collected hundreds of thousands of ultrasound pictures of the wrist’s nerves for the initial NerveTrack version. The doctors performing the annotation required a great deal of technological assistance, and the annotation approach depended on both private and open source tools created by and for AI experts.
The AI engineers at Samsung Medison realised that in order to create the most accurate models for the upcoming NerveTrack version, which would need new inference models for the nerve structures in the elbow, shoulder, and neck, they would need to gather tens of thousands of more annotated photos.
Due to the difficulties encountered in creating the first iteration of NerveTrack, Samsung Medison began looking into alternative approaches that would simplify and expedite the annotating procedure. The maker of medical equipment used Intel’s user-friendly, intuitive Intel Geti platform for the subsequent annotations.
In order to facilitate the annotation process for their second iteration of NerveTrack inference models, Samsung Medison was granted early access to the Intel Geti platform by Intel. As a proof of concept, Samsung Medison first tested the new annotation process provided by the Intel Geti platform on a small data set of just 13 ultrasound pictures. With its point-and-click graphical user interface and AI-enabled annotation assistance, the Intel Geti platform proved to be user-friendly and intuitive to a number of medical experts who took part in the successful proof of concept.
Annotation assistants on the Intel Geti platform make it simple for users to identify photos and draw designs. The platform offers annotation predictions following the annotation of a subset of photos. The labelling process is streamlined when the user approves or modifies the predictions as necessary. With minimal technical assistance, the physicians were able to swiftly and precisely annotate pictures of nerve systems in the elbow, shoulder, and neck to these user-friendly features.
The active learning feature of the Intel Geti platform reduces the quantity of annotations needed to train models and aids in automating data annotation. In their NerveTrack project, Samsung Medison used the active learning feature to provide AI annotation predictions, which resulted in a quicker and less time-consuming annotation procedure.
The image on the left represents ground truth, the target of the NerveTrack machine learning model. The image at right is annotated with the Intel Geti platform for inference model training proof of concept, based on 13 annotated ultrasound images.
Extending the Intel Geti Platform to Providers and Hospitals
Samsung Medison proceeded with the Intel Geti platform to gather tens of thousands of additional annotations in just two months after being pleased with the test findings.
The research received ultrasound video picture data from a number of South Korean hospitals. Physicians could work on their own and contribute their own knowledge and skill sets. The range of methods used helped lessen bias in the final data set after all the annotated photographs were submitted. It was anticipated that this would contribute to the new NerveTrack model’s increased accuracy and adaptability to a wider range of end users.
Intel Technologies Facilitate AI Deployment, Optimisation, and Inference
To develop the reference image database for the deep learning framework, Samsung Medison’s AI engineers and programmers sorted and processed the final data set of annotated images of nerve systems in the elbow, shoulder, and neck. The NerveTrack model’s AI inference engine and second phase were trained using that framework for those nerve locations. The model was tailored for the Intel Core i3 processor that powers NerveTrack on the intended ultrasound device using the Intel Distribution of OpenVINO toolkit.
Following regulatory agency approval, these new NerveTrack models will be implemented on Samsung Medison’s high-resolution ultrasound equipment with assistance from the Intel Distribution of OpenVINO toolkit.
In an eight-step procedure, Samsung Medison used a number of Intel technologies to train and implement the NerveTrack models for the elbow, shoulder, and neck:
- Obtain movies and pictures from ultrasounds
- Make a separate Intel Geti platform project for every bodily part.
- Hire medical professionals to manually annotate small sample data sets using the platform.
- Utilise the Intel Geti platform to train specific models.
- To annotate the remaining image data, use the active-learning capability of the Intel Geti platform for AI annotation predictions.
- Using a REST API, combine all project data sets from every hospital, then use the combined data set to train models.
- Use the Intel Distribution of OpenVINO toolkit to optimise and implement the models.
- Add the updated models to NerveTrack.
About Samsung Medison NerveTrack
Medical equipment manufacturer Samsung Medison, a division of Samsung Electronics, makes ultrasound devices. Samsung Medison is known worldwide for its innovative use of cutting-edge technologies and R&D. Doctors using Samsung Medison ultrasound scanners can use the company’s NerveTrack model, an AI-enabled inference model of human nerve systems, as a guidance for placing needles for local anaesthesia and other treatments.
