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 (b): 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.

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