Eigenvision is working in a research project together with researchers at Sahlgrenska
University Hospital and Skåne University hospital. The aim of the project is to develop
an automated system for quantification of skeletal tumor burden. Skeletal metastases
have high uptake in NaF PET images, but high uptake may also have other causes such as
wear or benign lesions. This means that to achieve an accurate quantification,
information on both the size and intensity of high-uptake regions as well as on its
anatomical location is required. Eigenvision is working on several subproblems required
to obtain this information.
Bone segmentation
The video on the left shows a whole-body CT image (computed tomography). This is the
input to our bone segmentation software. Using modern image analysis techniques, 49
different bones are detected and delineated, providing a segmentation result as
illustrated on the right.
Multi-organ segmentation
Similar techniques can be used to segment soft tissue. The
video shows a split view of the input CT volume and the automatic multi-organ
semantic instance segmentation result.
PET hotspot classification
The video shows an overlay of a positron emission tomography image (PET) on a CT image.
Regions with high PET uptake are delineated in white. Combining the information from the
PET image, the CT image and the bone segmentation, we aim to classify each individual
hotspot as either a benign lesion or a cancer metastasis.
In this project we are developing advanced tools for digital microscopy. This
includes free-form registration to account for tissue movement as well as stitching
errors from the microscope, but also automated cell detection and segmentation.
Together with Mometric, we are developing computer vision-based systems for
analysing competitive swimmer performance using structure-from-motion techniques and
markerless human pose tracking.
For example, by modelling refraction effects at the interfaces between air, glass and
water, perfectly stitched synthetic panoramas can be produced from multiple fixed
video cameras, covering the entire pool.
Using a combination of fiducial markers for absolute positioning and simultaneous
localization and mapping (SLAM), handheld cameras can also be used for measuring
swimmer position and velocity, as demonstrated in this video.
Client: Mometric AB
Induction welding control
Carbon fiber reinforced plastic, or CFRP, is an extremely strong and lightweight
material with applications in aerospace, automotive and sports equipment industry
etc. Together with Corebon, a leading company within manufacturing and handling
of CFRPs, Eigenvision has developed a new way of estimating the temperature in
the weld during induction welding of CFRP components.
Induction welding rig
Traditionally, the temperature is measured by placing a sensor, such as a
thermocouple, inside the weld. This complicates the welding setup and introduces
external materials into the weld, and it also wastes a temperature sensor that can't
be reused. In this project, we instead estimate the temperature of the weld using a
data-driven, deep learning approach. We train a model to output the weld temperature
based on signals that are available anyway, such as frequency, current, voltage and
phase shift.
To estimate the temperature we train a recurrent neural network, more specifically
an LSTM or Long Short-Term Memory model. Using a recurrent model enables processing
of variable length input and the model can be applied step-by-step in real time
during welding. In addition to estimating temperature, we also model the uncertainty
inherent to the data (aleatoric uncertainty) as well as model uncertainty
(epistemic uncertainty). An example of a trained model used to estimate weld
temperature (compared to ground truth from a thermocouple) can be seen above.
Client: Corebon AB
About
Eigenvision AB, founded in 2015, is a small company working with research and development in automated image analysis, computer vision and machine learning. We work as consultants automating your daily image analysis tasks, but also in long-term research projects. We have extensive experience taking algorithms from prototype to production code.
Olof Enqvist
Olof is Reader in applied mathematics with a specialization in image analysis. His research interests include geometric vision problems and semantic segmentation of medical images.
Sebastian Haner
Sebastian holds a PhD degree from Lund University in applied mathematics, specializing in computer vision. He focuses on 3D reconstruction, especially efficient real-time solutions.
Johannes Ulén
Johannes holds a PhD degree from Lund University in applied mathematics, specializing in computer vision. He is an experienced software developer with expertise in image segmentation, optimization and machine learning.
Måns Larsson
Måns holds a PhD degree from the Computer Vision group at Chalmers University of Technology. His research interests lie within semantic segmentation, with a focus on deep learning methods.
Magnus Linderoth
Magnus holds a PhD degree from Lund University in Automatic Control, specialized in robotics and computer vision. He is an experienced software developer with expertise in signal processing, real-time image processing and machine learning.
Daniel Bergh
Daniel holds a PhD degree in pure mathematics from Stockholm University. He also has a solid background in software engineering, with several years' experience as a developer and systems architect.
Carl Toft
Carl holds a PhD degree in Computer Vision from Chalmers University of Technology. His research mainly focuses on combining geometry, semantics and deep learning to solve problems in geometric computer vision.
Fredrik Kahl
Fredrik is an internationally renowned researcher and head of the Computer Vision and Image Analysis Group at Chalmers University of Technology.
Contact
We can help you with medical image analysis, computer vision and machine learning.