Skip to content

Latest commit

 

History

History
96 lines (72 loc) · 7.98 KB

README.md

File metadata and controls

96 lines (72 loc) · 7.98 KB
Raw

banner

Thomas Havlik's Medical AI Portfolio

I am a deep learning researcher that is applying for the 2024 medical school admission cycle. It seems beneficial to offer concrete examples of how I work with medical data. These projects are enumerated here, and each has its own page that details the challenges faced and the results achieved. A few other miscellaneous projects also made the cut.

In accordance with non-disclosure, none of my professional medical AI work is featured here.

Experiments

These experiments make use of ground truth provided with the data. Ground truth for medical data typically constitutes either empirical measurement or the judgment of attending physicians.

Completed projects

This is a list of tasks that I have "solved", meaning that I've trained a model capable of performing the task with respectable accuracy.

Works in progress

These are tasks that I haven't solved yet, mostly because training on fMRI scans is nontrivial and requires way more video memory than I have at my disposal. As the potential for machine interpretation of fMRI data is enormous, I am determined to deliver on a project with this modality.

  • StudyForrest (WIP): infer features of an audio-only film with fMRI
  • LA5c Study (WIP): predict participants' questionnaire answers from their MRI scans
  • Neural Rendering (WIP): single-shot, differentiable 3D rendering

Datasets

These are datasets that I have compiled. My interest in the modeling of visual manifolds was first inspired by recurrent dreams of mine that assume the form of video games, particularly those from my childhood. I have always found such media of dreaming to be both visually intuitive and semantically potent - attributes critical for unsupervised learning.

Tools

These are various pieces of software I have written that pertain to either medical education or machine learning research. I think the potential for software to enhance medical education is substantial, especially given the promises of machine learning.

  • e-Shadowing Transcriber: an application for enhancing the medical e-Shadowing experience
  • vpn-operator: a Kubernetes controller for VPN sidecars written in pure Rust
  • t4vd (Tool for Video Data): a collaborative, open source platform for easily creating labeled datasets from YouTube videos
  • Midas Download Tool: a collection of scripts that bulk downloads MRIs made public by Kitware on their now-retired MIDAS platform. The repository now serves as an archive for the difficult-to-find CASILab Brain MRI dataset comprised of healthy volunteers.

Running Code

Configurations are defined in .yaml files, which can be composed via the include: directive to conveniently form derivative experiments with minimal boilerplate. An experiment can be run by passing the path to the input yaml as the --config flag to src/main.py:

$ python3 src/main.py --config experiments/deeplesion/localization/basic.yaml

Note: the script assumes the current working directory is the root of this repository. By convention, all file and directory paths in yaml files are given relative to the repository root.

If an experiment hangs during the initial validation pass, it is likely because nonechucks is suppressing exceptions thrown by the dataset. This behavior improves fault tolerance, but can complicate debugging. To disable, set exp_params.data.safe: false in the experiment yaml.

Docker

Whenever possible, it is recommended to use Docker to ensure a reproduceable execution environment. Refer to Nvidia's guide on utilizing a GPU with Docker to enable the --gpus all flag:

$ docker build -t thavlik/machine-learning-portfolio:latest .
$ docker run \
    -it \
    --gpus all \
    -v /opt/data/dataset_name:/data/dataset_name \
    thavlik/machine-learning-portfolio:latest \
    python src/main.py \
        --config experiments/experiment_name/task_name/config_name.yaml

If you must run your experiments outside of a container, refer to the Dockerfile for notes on dependency installation. It's not possible to install the correct versions of some libraries using only pip install -r requirements.txt, hence those dependencies' versions are unconstrained.

Hardware Requirements

Almost all of these experiments are optimized to run on 11 GiB video cards and use as much memory as possible, either by increasing batch size or parameter count. Most experiments will not run on an 11 GiB card that is also driving a display, as the display itself typically requires at least a few hundred MiB.

Relevant Literature

Many of the ideas implemented in this repository were first detailed in the following papers:

  1. Auto-Encoding Variational Bayes
  2. 3FabRec: Fast Few-shot Face alignment by Reconstruction
  3. DARLA: Improving Zero-Shot Transfer in Reinforcement Learning
  4. Generative Adversarial Networks
  5. Progressive Growing of GANs for Improved Quality, Stability, and Variation
  6. beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework
  7. Towards Photographic Image Manipulation with Balanced Growing of Generative Autoencoders
  8. Understanding disentangling in β-VAE
  9. Analyzing and Improving the Image Quality of StyleGAN
  10. Instance Normalization: The Missing Ingredient for Fast Stylization
  11. Deep Residual Learning for Image Recognition

Contributing

This repository was intended to be repurposed. As part of the open source community, I do not have the perception that a minor contribution or bug fix from someone else dilutes the claim that this repository is representative of my capabilities. Issues and pull requests are welcome.

License

All code in this repository is released under MIT / Apache 2.0 dual license, which is extremely permissive. Please open an issue if somehow these terms are insufficient.

Contact

thavlik (at) protonmail [dot] com