Diagnostics in histopathology is still based on slow, and often subjective, manual approaches.
One emerging technology that can provide a solution to this problem is machine learning or, specifically, deep learning. We believe that, by automating routine tasks and quantifying various diagnostic parameters, deep learning algorithms will lead to one of the biggest transformations in pathology.
Unfortunately, despite its potential, the adoption of these technologies into routine practice has been slow. There are two main barriers to the widespread use of deep learning in veterinary pathology:
1. Lack of infrastructure required for the digitization of histological specimens.
Due to their high cost, many institutions and clinics still can’t afford microscope slide scanners and therefore are not able to make a transition from traditional microscopy to digital pathology. In addition to their high price, these systems are often completely closed and use proprietary technologies and image formats.
These practices are slowing down science development significantly. Digitization is crucial because it “opens a door” for the use of computer algorithms, and therefore, deep learning in image analysis. Without a high-resolution and inexpensive microscope slide scanner, we can’t expect widespread use of these algorithms in digital pathology, regardless of their potential.
2. Low-cost and scalable solution for creating Deep Learning applications in veterinary histopathology.
Besides digitization, deep learning algorithms need a lot of labeled images. Labeling medical images is not only a time-consuming task, but it also requires domain expertise. Domain expertise coupled with time leads to high costs. High costs are not sustainable, especially in veterinary medicine. Also, having machine learning experts training neural networks without a deep understanding of the underlying data is very inefficient, and sometimes not so harmless.
Instead of looking at these barriers as two separate problems, we should take a more “holistic” approach to these issues. We argue that the most efficient and scalable way to create hundreds of deep learning applications we need in veterinary histopathology is to build a strong community that will have the skills and intrinsic motivation to do the whole end-to-end process needed for creating deep learning models.
In order to build that community, we designed Marvin — a low-cost, open, and programmable microscope slide scanner. For veterinary students, Marvin has a similar purpose to the Raspberry Pi. As programmable hardware, it serves as a great introduction to python programming, but also as a stepping stone into more complex image processing. As a microscope slide scanner, it is a platform for developing and deploying their models.
Ahead of us
Our goal is to build a platform consisting of hardware (Marvin) and software (cloud) components that will enable community (veterinary pathologists) easier development and deployment of their deep learning applications.