Imorphics have developed revolutionary patented technology to radically improve the performance of 3D Statistical Models to automate segmentation and analysis of virtually any anatomical structure or tissue in a 3D medical image.
Our statistical learning technology has been developed for over 15 years and is based on what has been proven to be one of the most successful methodologies for medical image segmentation. The technology uses a set of learning examples that represent variability of an object’s shape and appearance to train a flexible anatomical model before the shape and appearance of this model is altered to best match an object seen in an image.
This advanced trainable platform demonstrates fully-automated identification and segmentation of tissues with sub-millimeter accuracy, which has been validated on anatomy including:
- bones, cartilage, meniscus and other muscloskeletal tissues;
- sub-cortical brain structures;
- prostate, liver, kidneys, aorta and lungs
- salivary glands, optical structures and sinuses.
Because the model being used can only deform within the limits of what was previously learnt, this prevents the model from taking up implausible shapes when representing an object in an image. The result is a very robust segmentation solution for noisy images, incomplete anatomy or images of low contrast.
Using Imorphics machine vision technology, we have demonstrated fully-automated identification and segmentation of tissues such as bones, cartilage and other musculoskeletal tissues, sub-cortical brain tissues, prostate, liver, and other abdominal organs, skulls and sinuses with sub-voxel or sub-millimeter accuracy.
In segmentation performance, Imorphics have won all four of the MICCAI Grand Challenge segmentation competitions that they have entered:
2010: SK10 Knees Grand Challenge;
2012: PROMISE12 Prostate Grand Challenge;*
2014: VISCERAL 2014 Thorax/Abdominal Grand Challenge;*
* Imorphics had done no previous work on either prostate or visceral images before developing winning solutions in less than two months.
In contrast with traditional approaches to image analysis, our statistical modelling technology outputs a dense set of real anatomical landmarks when a model is matched to an object in an image. This means that the shape and appearance of anatomical objects can be properly and rigorously compared across time points and human subjects. This enables the careful comparison of populations; identifying significant differences within a population as the disease progresses, or identifying shape differences based on gender, ethnicity, size and phenotype.