Medical Image Synthesis
Our client was an organically growing start-up, which was already having a first generation product in the medical space, hence their product-market fit was validated, however, the existing product required too much manual tuning on the user side. They were looking to develop a second generation product with in principal higher quality and also more automated functionality. It was also a requirement to meet certain touch points so that manual editing still remains a possibility though.
TechnoLynx has started working on the basis of one of the state-of-the-art models in agreement with the customer, and at first we only made cautious modifications in order to meet functional requirements for synthesis. After having satisfactory results relatively early on in the project, the scope has been expanded so that it covers the entire end-to-end image processing pipeline, including segmentation, classification and structural deformation, not just synthesis alone.
Once the initial functionality was put in place for all parts, we executed a series of experiments, some based on ideas of other state-of-the-art models, and some were genuinely new ideas, previously not used in this context. This was the foundation of an iterative improvement model that we could use for the management of the project enabling us to deliver gradual improvements to parts of the pipeline in an on-demand basis.
We used PyTorch for this project, and we still believe it is a good starting point, especially if it is expected that a great deal of research shall be involved. However, it must be pointed out that during delivery the use of PyTorch probably posed more challenges than a reasonably optimized delivery with TensorFlow.
A functionally complete end-to-end pipeline has been delivered as well as clear yet flexible roadmap and constant flux of improvement ideas, based on novel research ideas.
Additionally to the successful delivery of the required functionality and quality, TechnoLynx has also assisted the client with its attempt to protect the novel IP that has been created as part of this R&D project.