Author: McKell Woodland

Poster entitled “StyleGAN2-based Out-of-Distribution Detection in Computed Tomography” will be presented at the Practical Big Data Workshop (Ann Arbor, Michigan).

Abstract entitled “Comparing Transfer Learning, Data Augmentation, and Data Expansion in the Improvement of Medical Image Generation” will be presented in the 2022 AAPM annual meeting during the “AI/ML Autoplanning, Autosegmentation, and Image Processing 1” session. This session is from 1-2pm in Ballroom B.

Teaching CSE 450: Machine Learning & Data Mining

Abstract entitled “Improving the Generation of Synthetic Medical Images using Data Augmentation and Transfer Learning” chosen for a poster presentation at the Young Investigator’s Symposium at SWAAPM 2022. The top 20 abstracts were chosen for the session.

In addition, my presentation entitled “Houston, Our AI Models Have a Problem!” was selected for the MedPhys Slam competition.

I was the lead instructor for 18-week coding courses for elementary and middle school students in the Glendale Unified School District. The middle school course had 200 students, while the elementary school course had 100. Both courses were taught through Zoom. I taught computer science basics, including: if/else statements, for loops, while loops, and functions through the use of Python Turtle. The course was administered by codeConnects, a program sponsored through The Coding School.

Held office hours for undergraduate students participating in the 2022 Rice Datathon. Mentored students in basic data manipulation, data visualization, machine learning models, and time series analysis.

Award given at the Practical Big Data Workshop 2021 put on by the American Association of Physicists in Medicine.

The paper is entitled “Correlation of In-Vivo Imaging with Histopathology: A Review”.

https://doi.org/10.1016/j.ejrad.2021.109964

Poster accepted into the Practical Big Data Workshop 2021 by AAPM.

Detecting Out-of-Distribution Images for Active Learning using a Generative Adversarial Network (StyleGAN2)

Purpose: To detect outliers from a medical imaging distribution used to train deep learning models. The proposed method can identify Computed Tomography (CT) scans that would make a clinically deployed segmentation model fail.

Methods: A cohort of 143 patients with 147 non-contrast enhanced abdominal CTs (97 training, 50 test) was used. A StyleGAN2 network, a state-of-the-art high-resolution generative model that uses a style-based generator and backpropagation to encode, was trained to reconstruct slices. Data was preprocessed with windowing, masking, and conversion to 512×512 PNG images. The network’s generative quality was measured with the Fréchet Inception (FID) and Wasserstein (WD) distances. Slice reconstructions from the test CTs with a learned perceptual image patch similarity score (compares VGG network feature representations) over 0.1 were classified as out-of-distribution.

Results: Randomly generated slices had FID and WD metric values of 8.77 and 0.10, respectively. All test images on which a segmentation model failed (the model had a Dice coefficient of 0.96 on test CTs) were classified as out-of-distribution.

Conclusion: A paradigm was optimized to predict when a clinically deployed segmentation model would fail with a 100% success rate. The paradigm could be further used to create heterogenous imaging datasets and prioritize images for labeling.

• Advised a student who built a glioblastoma detector on MRI scans using a pre-trained CNN
• Guided another student who built an arrythmia detector on ECG signals using a pre-trained ResNet
• Gave a talk on deep learning in medical imaging to ~300 high school students