Explainable Disease Classification via Multi-Ultrasound Images

多張超音波影像之可解釋疾病分類

By: Ian Liu 劉以恆^{1 2 3}

Project Mentor/Advisor: Tso-Jung Yen 顏佐榕, PhD³

Workflow

Example Results

Fig (a): correlation coefficients of each image, indicating marginal influence of each image.

Fig (b): ElasticNet coefficients of each image, indicating conditional influence of each image. Faded bars indicate statistical insignificance.

Fig (c): Ridge coefficients of each image, indicating conditional influence of each image. Faded bars indicate statistical insignificance.

What Makes Our Study Different

• Traditional LIME is only applicable on single input (ex. single image). We extend LIME to graph neural networks (GNN) by applying principles of LIME on nodes and edges of a graph neural network.
  • Instead of randomly perturbing "superpixels" (segmentations) and creating variations of the original image, we use graph sampling to create variations of the graph and create local models from the subgraphs.
  • This allows us to derive image-level importance and influence for each subject.
• LIME uses traditional local regression/classification, so it can only display conditional relationships. For example, typical regression interpretation of coefficients is: "given other variables do not change, so and so variable has such impact." We display marginal relationships by calculating correlation.
• Our approach makes use of summary statisics such as confidence interval and standard errors, which allows for uncertainty quantification.
• We employ a novel two-stage adaptive class-balanced sampling method to encourage class balanced samples.

Order of Running the Pipeline

1. generate_samples_and_marginal_relations.py (which imports from marginal_relation.py)
2. ridge_run.py OR/AND elastic_net_run.py (which both import from classifiers.py)
3. organize_output.py

Explanation of Each Step

1. Generate samples to create explanations on. The sampling algorithm tries to create class-balanced samples. Then, the correlation is calculated for each image with the model's predictions.
2. Ridge/Elastic-Net logistic regression to identify important coefficients.
3. Organize output into a folder structure as such:

• organized_output_folder\
  • correlations
    • positive
    • negative
    • neutral
  • csv
    • 0-elastic_net_coefficients.csv
    • 1-ridge_coefficients.csv
    • 2-correlations.csv
  • ridge
    • positive
    • negative
    • neutral
  • elastic
    • positive
    • negative
    • neutral
  • plots
    • 0-vbar.png
    • 1-hbar.png

Here, each positive folder contains original images that contribute to model predicting the positive class, while the negative folder contains original images that contribute to model predicting the negative class. neutral folder contains non-significant images. plots contain all bar plots, while csv contains the coefficients for each image.

Footnotes

1. Department of Data Science, Fei Tian College Middletown, Middletown NY ↩

2. Department of Biostatistics, Brown University, Providence RI ↩

3. Institute of Statistical Science, Academia Sinica, Taiwan ↩ ↩²