Pathology Imaging Segmentation
Neural Networks PyTorch Python React DockerImproved segmentation-based methods for automating Cobb angle measurement in Adolescent Idiopathic Scoliosis by analyzing the effects of training hyperparameters and data augmentation techniques, achieving a SMAPE of 10.30% with SalsaNext, surpassing reference models
Streamlined Training and Evaluation Pipelines
There were more than 40 models we needed to train within a limited time. And many of them would take hours to train on an NVIDIA A100 GPU instance.
The initial manual processes involved were cumbersome and inefficient, largely due to the necessity of manual data handling within and between steps:
- preparing and dividing datasets
- training models
- conducting segmentation inference
- calculating SMAPE scores from inference results
I created two pipelines that automate these processes
-
Training Pipeline
- automated the queuing of training tasks on multiple GPU instances for parallel training sessions
- maximized our computation resource utilization
- cut down on project time significantly
-
Evaluation Pipeline
- automated the benchmarking of trained models. This system streamlines the evaluation process
- produced direct visualization of prediction accuracy and training details via a web app
Custom Data Loader Development
The models we tested require different input image size and channels
I developed a custom data loader that can load all of our training data from a single source and serve multiple data formats and image sizes based on the model's requirements
Dashboard - Training Result Visualization
After a model goes through the evaluation pipeline, the benchmark of the model will be automatically generated as a web page. This allows researchers to easily reference these results when writing reports.
Here are some examples of the benchmark page of a specific model:
Explore model performance visualizations in your browser: (hosted in Docker)