To mark the recent launch of Patholytix AI 1.5, we describe how Patholyix AI fits in the technology landscape available to AI developers working on Pathology AI projects, and how the platform can benefit them in accelerating impactful data generation for their pathology stakeholders.
Pharma is investing heavily in exploring the potential of AI to support the acceleration of drug development processes. In Pathology, AI has been identified as a considerable productivity driver, facilitating high throughput screening of content and generation of novel data, which to date is unavailable to scientists. Data scientists working in Pharma and Biotech have considerable challenges in gathering, annotating and cleaning pathology data used for AI model development. These data scientists are reliant on multiple stakeholders within their companies to contribute to the data generation process, many of these stakeholders are often located in different physical locations within the organisation.
Leveraging AI in pathology, initially requires data to be digitised using slide scanners for whole slide imaging, the images then need to be reviewed, diagnosed and annotated accurately to facilitate their use in AI model development. AI engineers and Data Scientists are involved in curation of these datasets and are also involved in coordinating annotation processes across these stakeholders and research domains. Many different pharma organizations use different whole slide imaging scanners and digital pathology application software between different research domains. As a result, different tools are used to review and annotate specimens, with variant annotation formats, such as vector-based or pixel-based annotations. In general, annotation tools currently available in the market do not allow for the highly scalable and high throughput annotation requirements for AI research. To support this process researchers are often developing their own tools which can be time-consuming and challenging to maintain as the digital pathology technology landscape evolves.
There is currently no standardisation between scanning devices. The optical appearance and pixel resolution provided by the scanners can vary from instrument to instrument. The difference between scanners can change the visual appearance of the image - If the different scanners haven't been calibrated in some way to guarantee they have a constant visual appearance then you will get variation not only between scanners but within data produced by a single scanner. This generates an additional challenge to enable resultant models to generalise between datasets generated on different devices.
The overheads of developing a pipeline facilitating the integration of data from across the pathology network, along with the complexity of dealing with the variety of available scanning devices creates a considerable technical overhead for AI engineers and Data scientists, before ever getting to focus on the model architectures available for problem solution.
If an AI engineer wants to test a variety of different models and parameterisations of models they will need GPU clusters to run, train, test and validate the algorithms. Depending on available hardware, experiments can take hours, days or even weeks. If the AI engineer would like to run several of these experiments in parallel, they will need access to a grid of AI engines, running across multiple GPUs, to facilitate this. Orchestration of these runs on a GPU cluster can create significant additional overheads for the development team.
Patholytix AI seamlessly integrates the contributions of pathologists and AI engineers, connecting them in a coherent way to work much more efficiently and effectively together. Patholytix AI solves bottlenecks in data management and model development by providing all of the necessary tools for selecting images from studies of interest, a mechanism to rapidly and ergonomically annotate those images and a full AI pipeline that takes generated annotations and associated image data and provides it to the requisite models for training in a standardised way. Using our pipeline, AI development is easier, faster and more efficient, allowing AI engineers to focus on generating high performing models.
Patholytix AI allows the AI engineer to easily standardise input data from multiple scanning devices meaning everything within the pipeline is consistent and enables a common framework for model training and development. To allow for the generation of robust models, several integrated data augmentation strategies are provided, such as class balancing, elastic deformation, colour and geometrical data augmentation. A multi-magnification AI pipeline is also provided to maximise contextual understanding of the problem being resolved.
AI engineers invest a lot of time into researching state of the art models - Patholytix AI mitigates this by having a number of already integrated high performing model architectures, that are state of the art in pathology research. Our team is continuously evaluating various models in literature and making the best available to our clients through incremental releases. As part of Patholytix AI 1.5, we have provided clients with a new model architecture that allows the user to combine different state of the art backbones and encoder-decoder structures. Patholytix also provides a mechanism where the AI engineer can take another algorithm and plug and play into our architecture allowing easy integration of your own model into the pipeline. If you have experience with Keras and Tensorflow then you can write your own model and plug and play with our analysis pipeline.
Patholytix AI is also highly scalable in line with your requirements. With Patholytix AI 1.5, we introduce the ability to orchestrate either Linux or Windows GPU clusters. Using the benefits of Docker Containerisation, you can run many engine instances concurrently on one Linux platform allowing you to run many model configurations in parallel. If running larger models, leverage the potential of the full GPU cluster concurrently to train a single model as required.
We provide a number of integrated performance validation strategies for the quantitation of model performance and allow you to share whole slide inferences from developed models with your Pathologist stakeholders for review. These whole slide inferences work seamlessly in our Patholytix content management solution giving your pathologists a good visual sense of the performance of your models.
Finally with Patholytix AI, you also get the benefit of hardened production code that has been put through a rigorous validation process, continuous evolution of the platform and improvement of capabilities along with a strong roadmap of planned innovations.
To learn more about Patholytix AI please contact firstname.lastname@example.org to arrange a consultation with our AI specialists.
Invent, DCU, Dublin 9, Ireland