While working with a variety of startups on computer vision projects, I noticed a pattern: models often don’t generalize when you expect them to. And, unfortunately, this can come as a surprise late in the project. I quickly learned to
In deciding which approaches to try, I learned to look at it through the lens of generalizability.
Now I help startups do the same. Given the unique characteristics of a pathology dataset, I help them identify the best path to models that handle the expected variations, enabling them to take their products to market sooner.
Our team was fortunate to work with Heather on a short-term project, where she was a critical contributor to establishing a framework for future studies. She has an unusual ability to demystify, in plain language, new developments in computer science for a wide audience, and I am confident through her work she will influence many teams and help bring to reality the next set of tools to transform pathology.
Keith Wharton Ultivue, VP Medical Director
Working with Heather and Pixel Scientia Labs was a great choice for us because of her extensive experience in cancer detection, imaging, and machine learning. Based on this background and expertise, she was able to support the team as we worked to empower our imaging solutions through AI and machine learning.
Alan Kersey CytoVeris, President & CEO
Heather’s knowledge of the current state of the art within the digital pathology field is second to none. Discussions of prior art enabled our team to focus on novel research and refine our current AI methodologies for clinical research.
Jenny Fitzgerald Deciphex, Director of Clinical Research & Operations
Large & Diverse Images
Whole slide histopathology images are often more than 60,000 pixels across. They contain multiple tissue types and both tumor and non-tumor tissue. Tissue appearance is heterogeneous, both from patient to patient and sometimes within a single tumor.
Weak Labels
For some applications like mitosis detection and tissue segmentation, pathologists can provide detailed annotations. But for patient-level prediction tasks like molecular biomarkers, treatment response, or patient outcomes, the algorithm must learn itself which regions of the image are important, often employing multiple instance learning.
Limited Labeled Data
Some applications of computer vision make use of millions of labeled images, whereas data sets of 1000 or so patients are much more common for medical applications. Transfer learning and self-supervised methods are often critical to success in this low sample size regime.
Additional Modalities of Data
The most powerful models that can improve patient care and outcomes often use multiple modalities of data - histopathology, clinical, genomic, proteomic, etc. Specialized models can make use of these structured and unstructured sources of data.
Language Barrier Between Disciplines
Understanding the intricacies of a particular application and possible clinical use cases requires the expertise of pathologists and other domain experts. Project success depends on communicating both ways - about the disease and about the machine learning solution.
The process for machine learning is empirical and iterative - hypothesize a model, test it, and improve it.
While many talented machine learning engineers can create and train a model, they may be inexperienced with the challenges posed by real world data. The complexities of massive images and noisy or missing labels are a whole different ball game than clean benchmark data sets.
Machine learning engineers may also lack the experience to identify unique aspects of data that, with a customized model, can improve predictions.
I spent my Ph.D. developing solutions to study breast cancer and learned to create new machine learning methodologies motivated by particular aspects of the research data.
The same may be beneficial for your project, but you won’t know it without looking from the right perspective.
