Nobleo developed an algorithm for the segmentation of structures in ultrasound image video streams. Ultrasound imaging is typically known from echography for unborn babies. Due to the low cost, small size, and non-invasiveness, it is commonly used for many other diagnoses, such as measuring bladder volume, cardiac inspections, analysis of blood flows in arteries and veins or diagnostics of muscle fractures.
Ultrasound images have a noisy nature, and even for experts it is hard to derive information from a single image. Therefore we have created a time domain based solution. Multiple novel techniques have been applied like recycling information from a previous image to obtain a more accurate segmentation of the current image.
The algorithm tracks ultrasound features in a sequence of images using a Long- and Short-Term Memory (LSTM) network in combination with mask tracking. The algorithm allows for tracking of ultrasound features as well as automatic selection of a frame in a sequence.
Next to the above, novel labelling techniques were used to avoid spending time and resources on unnecessary, laborious image annotation jobs. In addition, we designed our machine learning based algorithms to only focus on those aspects of the data that are relevant, similarly to how a human would interpret the data. This way we can quickly derive new information from images that can guide a medical domain expert to a more accurate or confident observation.
Please reach out to us if you are looking for: