University of Pittsburgh

Innovation

Practical Applications Solving Real Problems

Established in January 2024, CPACE has already become a recognized national and global leader in health care AI, notably within pathology and laboratory medicine.

The unique tools developed within CPACE are now being used to enhance diagnosis, patient management, reporting and decision-making along with workflow efficiencies.

We have launched multiple practical AI tools that include Report-Genie, Auto-Pix-AI, WSI-Genie, Pitt-GPT-Plus, and Nebulon-GPT which are starting to transform and enhance our diagnosis, reporting and decision-making domains. Several of these are under patent review.

Our hybrid approach, combining vendor partners and home-brewed tools, is creating a playground that is expediting our research, quality and educational endeavors. For our own GPTs, their specialness comes from the fact that they’re 100% private, because they’re running 100% on our own neural networks.

“We are not just hypothetical. We’re very practical. We actually know how to make the AI because we live the problems. They’re in our own department. We don’t have to go and ask people to annotate for us or explain something to us.”
– Liron Pantanowitz, Chair of the Department of Pathology

CPACE Innovations/Products

Pitt-GPT-Plus

As opposed to paying vendors to use platforms that consume 10 times the energy of the neural networks that we’re using to ask very basic questions, CPACE relies on an open source neural network with one-tenth or one-hundredth of the energy cost. Work that can be done internally without a loss of accuracy can be much cheaper, much faster, much more private and more sustainable.

Instead of bringing in the hardcore, heavy-duty neural networks, we can do the same job within this framework with multiple agents cross-checking each other with a fraction of the energy use.

That offers flexibility in cases where the most powerful models on the planet, with their high energy demands, are not needed.

Advantages include:

  • Reduced hallucinations: domain-specific knowledge integration
  • Higher accuracy in related queries
  • Cost-effectiveness: reduced reliance on expensive API calls
  • Scalability for institutional use
  • Enhanced control
  • Improved data governance
  • Enhanced data security (local, with no cloud)
  • Easy customization (flexible to solve any problem)

Other CPACE Innovations/Products

  • WSI Genie
    Whole slide image automated machine learning framework
  • Nebulon-GPT
    Completely private and on-prem ChatGPT-like application
  • AutoPix-AI
    Flat file image automated machine learning framework
  • Report Genie
    A combined rule-based and large language model integrated diagnostic report generator
  • Letter Genie
    A multi-agentic GenAI-based recommendation letter generator
  • And many more…

Vendor Applications

CPACE also uses several AI tools (MILO, STNG, V7, Halo) to enhance diagnosis, reporting and decision-making.

MILO-ML (Machine Intelligence Learning Optimizer)

Developed at University of California, Davis, by Rashidi et al.

MILO-ML is an advanced set of data science applications (seven apps under one umbrella) that includes its powerful automated tool for binary classification, designed to meet the rigorous demands of academic research while providing actionable insights from data. It offers:

  • Fully automated, high-fidelity binary classification
  • Secure on-premises deployment
  • Optimization for small to medium cohorts
  • An algorithm-agnostic predictor
  • Adherence to the CRISP-DM methodology
  • Incorporation of the top seven binary classification algorithms
  • A comprehensive preprocessing suite of data science apps

For more information:
https://milo-ml.com
https://www.nature.com/articles/s41598-020-69433-w

STNG (Synthetic Tabular Neural Generator)

Developed at the Cleveland Clinic by Rashidi et al.

Health care data accessibility for machine learning is encumbered by a range of stringent regulations and limitations. Using synthetic data that mirror the underlying properties in the real data is a solution to overcome these barriers. STNG comprises multiple synthetic data generators and integrates an automated machine learning module to validate and comprehensively compare the synthetic datasets generated from different approaches.

For more information:
https://www.nature.com/articles/s41598-024-73608-0

V7 (Automated Image Segmentation Tool)

V7 allows for handling massive medical imaging files and supports large whole slide imaging files, SVS format and multimodal data. Dynamic rendering for pathology samples provides for adding thousands of annotations at native resolution and the highest quality.

For more information:
https://v7labs.com