Artificial Intelligence, or AI, has become immensely prevalent in our daily lives. Usages of AI include transcribing speech into text for finding locations in Google Maps, translating languages, or extracting text from images through Optical Character Recognition (OCR) when looking at foreign menus. When it comes to the medical field, AI is also highly significant, being implemented in a variety of ways, from 1) Classifying and identifying diseases from medical images 2) Segmenting medical images, such as selecting regions in the image that are potentially tumorous 3) Detecting medical substances, such as abnormal cells from microscopic images, as well as other usages that has allowed AI to play an important role in medicine.
In this passage, we will discuss further about the use of medical image segmentation, the existing models that are used in the industry, as well as the AI research that Meticuly has applied to our titanium bone implant design.
Image Segmentation is a technique used in image processing to isolate a Region of Interest (ROI) within a picture, utilising pixel-level data to help determine this ROI. This technology has been extensively employed in the aiding of diagnosing abnormalities from medical images, including cell segmentation from microscopic images, abnormal tissue detection from ultrasound or X-ray images, as well as brain segmentation for grey matter, white matter, or the skull from CT scans or MRI.
Initial research on medical image segmentation focused on utilising image features such as edges or colour values to identify ROI. However, contemporary models that have exhibited great performance and experienced wide adoption are neural network models. Popular neural network models include the U-Net model, which operates based on Convolutional Neural Network (CNN) fundamentals, as well as further developments like the UNETR model which operates based on the Transformers model. These models are able to undergo effective training, by establishing relationships within the input image dataset.
At present, there are several tools available to simplify the training process of image segmentation models, including proprietary softwares and open-source libraries. One popular open-source library widely used in medical image processing is the Medical Open Network for Artificial Intelligence (MONAI) that has pre-trained image segmentation models that can be readily applied to desired datasets.
Furthermore, in April 2023, Meta announced the Segment Anything Model (SAM) that can segment images simply by clicking on objects within them. Researchers have started using SAM in the medical field under the name of MedSAM, and have found out that MedSAM can efficiently collaborate with medical professionals and segment a diverse range of medical image types.
In the future, it is expected that medical image segmentation and data collection tools will be more widespread in the medical field, and may replace traditional data collection methods known to be time-consuming.
Meticuly’s Application of Medical Image Segmentation
Indeed, the Meticuly team has leveraged medical image segmentation methods to aid the engineering team. The use of AI models has streamlined the design process, enabling the team to produce titanium implants for its users at a greater efficiency and speed.
A prime example of Meticuly’s application is its utilisation of image segmentation models to extract cranial features from 3D CT scans, in aiding the design of cranial implants. The integration of these models has streamlined the process for isolating cranial features, thus increasing the speed of communication between the team and medical professionals. Moreover, other models, such as generative models, can be used to predict outcomes for the implants. Originally, these processes would have taken 6-10 hours, but with AI models, the team can now transform 3D CT scans into rough 3D models within 5 minutes. This efficiency greatly helps surgeons’ surgery planning and has enhanced communication between Meticuly’s engineers and medical professionals.
Written by Titipat Achakulvisut
Translated by Eclair Sakdibhornssup
Comments