Comparing ULIME with L2 LIME and Cosine LIME

• To compare L2 LIME, Cosine LIME and ULIME on a limited imagenet dataset, please open the following Jupyter Notebook in Google Colaboratory.
• Instructions on how to run the code are given in the notebook.
• The code has also been sufficiently commented. If something remains unclear, please reach out to minhaj3737@gmail.com.
• Make sure you set the runtime in Colaboratory to GPU before running the notebook! https://colab.research.google.com/github/ansariminhaj/ulime_github/blob/main/Imagenet_github.ipynb

Training a CNN model on Liver and Pancreatic Tumors

• The dataset has been kept private due to patient data confidentiality. However, the outputs from our actual experiments have been preserved.
• Both liver_model.ipynb and pancreas_model.ipynb show the train/validation loss graphs and performance metrics of the trained models.

Domain Specific Superpixelization

• The dataset has been kept private due to patient data confidentiality. However, the outputs from our actual experiments have been preserved.
• domain_superpixelization.ipynb shows our superpixelization method applied on liver and pancreatic tumors.

Here are some of our domain specific superpixelization ULIME explanations for Liver and Pancreatic tumors. We first compare our version with the authors default quickshift algorithm.

We then request comments from a professional radiologist (Dr. Richard Kinh Gian Do from the New York Sloan Kettering Cancer Research Center), verifying the explanations.