Case Studies

Multi-GPU cloud training of 3D segmentation AI

A client needed an AI to perform 3D segmentation of brain tumors from MRI scans.

While they had some in-house experience with algorithm development, they did not have expertise in renting clusters of cloud GPUs, training an AI neural network model on more than one GPU at a time, checkpointing, hyperparameter optimization, or creating intuitive real-time dashboards.

Theta Tetch AI built the client’s pipeline and trained a bleeding-edge transformer-based 3D segmentation model on a cluster of eight NVIDIA A100 GPUs all training the same neural network.

We set up web dashboards for the client to watch the segmentation results live and monitor the training accuracy in real-time.

AI Tech

  • Neural Network: 3D Swin-Unet Transformer
  • Multi-GPU training in cloud using Torch Lightning framework

Data Analyzed

  • 3D, volumetric brain MRI scans
  • Multi-modal data (T2-w, T1-w, T1+Contrast, FLAIR scans)
  • 3D volume meshes of tumors

Strategy & Execution

  • Gathered a variety of MRI scans from multiple institutions
  • Visualizing data and segmentation results in 3D
  • Preprocessing (e.g. histogram equalization) to normalize scans
  • Setting up cross-validation experiments
  • Reporting Dice accuracy on Weights & Biases dashboard during live training
  • Deployed model in Docker container
Theta Tech developed a 3D volume overlay on an MRI scan.
Colors represent the probability of cancer development.

Cloud webapp for AI clinical study for doctors

A client had developed several in-house algorithms for helping diagnose Crohn’s disease but had no clear way to share their results with doctors and recruit doctors to be part of a clinical study on evaluating the algorithms.

The client’s code was not in a great state, and they had no experience building professional, user-friendly software for doctors.

Theta Tech AI first fully understood their needs, the clinical problem, and the algorithms.

Then, a serverless cloud-based web app was built to allow radiologists to view the algorithm results and click buttons to evaluate the algorithm’s performance.

Everything was logged to a cloud database for later statistical analysis by the client.

‍

AI Tech

  • XGBoost combined features to provide probability of isease
  • Compared doctor performance with and without AI

Technology Used

  • Customized OHIF viewer
  • Different logins for different doctors in study
  • Postgresql database for storing clinical study results

Strategy & Execution

  • Focus on user experience to make webapp intuitiv for doctors
  • Biostatistics to analyze doctors' results versus AI results
  • Visualization of AI results directly in radiology viewer

Data science for a medical device startup

A startup had built a proprietary device for measuring brain waves of sleep apnea patients.
With a deluge of data, they did not have the in-house data science expertise to help find patterns, extract business value, and design AI algorithms around it.

Theta Tech AI visualized the client’s data, analyzed it, and understood the business case and clinical problem.

Then, an XGBoost algorithm was implemented to predict episodes of sleep apnea using the patient’s brain waves before the episode occurred so the device could provide automatic intervention.

The client did not want to hire a full-time data scientist to extract business value from their data, but it still needed to be done.

They required an AI company that understood the engineering and had experience with healthcare problems and business value.

Algorithm for collaboration & distribution

A client had several in-house algorithm prototypes written in MATLAB and several algorithms from published papers they wanted implemented in Python.

However, their in-house developers did not have experience converting code from one programming language to another or implementing algorithms from publications.

Theta Tech AI converted the algorithms (whether prototypes or publications) to high-quality, reusable Python code.

Libraries were built with pip and hosted on PyPI so collaborators, vendors, colleagues, and researchers immediately use the client’s algorithms.

Theta Tech AI designed cloud Jupyter notebooks demo-ing the algorithms, thus allowing the client to distribute and share their work efficiently.