NSF MRI: High Performance Digital Pathology Using Big Data and Machine Learning


The overarching goal of the project was threefold. First, we integrated digital imaging into Temple University Hospital’s clinical operations. Then, we established a public database of pathology slides. Lastly, we developed an image classification system to automatically annotate each pathology slide. The result was a fully functional digital pathology imaging system. After the database of pathology slides were annotated using this system, they were archived and used for support, education, and/or research purposes.


Pathology slides are scanned with the Leica Aperio AT2 Digital Whole Slide Scanner (DWSS). The Aperio AT2 has a 400-slide capacity brightfield scanner allowing a high throughput rate of approximately 50 slides per hour at 20x magnification. The scanning software uses lossless compression and generates high quality TIFF files ranging in size from 250 megabytes to 5 gigabytes (resolution dependent).

Our App Server is what hosts the Leica Aperio eSlide Manager and Image Scope applications. The eSlide Manager is a web based application that pathologists can use to view digitized pathology slides and export specimen data to Image Scope; which is simply a local application.

To store the slides, we used a cost-effective Petabyte server. There are two Petabyte machines in total, one to store the digital slides, and one as backup. Since the digital pathology slides we use can sometimes have a size of up to 5 gigabytes each, we needed as much storage as we could get. Our Petabyte machine collected a 10-year digitized archive of digital pathology slides. Each Petabyte machine contains two JBODs with each containing 60 hard drives each. This means that there are 240 total hard drives installed on both machines combined.

Data Development

Our department has a successful history in managing medical databases. The TUH EEG Corpus now exceeds over 50,000 EEGs spanning the years 2002 to present. For this project, we have digitized a 10-year archive of existing slides and integrated digitization into ongoing clinical operations. Our student researchers worked side-by-side with clinical and research pathologists to develop best practices for system optimization. After a baseline system was established, the data was released for review. We collected feedback from users with respect to the scan parameters and organization of the public data to maximize the value of the data for the community.

Algorithm Development

We adapted our existing TUH EEG Corpus deep learning system to this project. Once a baseline system was in place, we compared our analysis with physician feedback and adjusted the system accordingly. The software was developed according to strict design and formatting requirements and included regression testing and built-in certifications. Our software was developed using Python and executes on general purpose processors and high-speed GPUs. As always, the data and software we developed is open-source.