The Importance of explainable AI Digital Pathology.pptx

This is a presentation about XAI in Pathological Image Processing

1. Cell level feature and function analysis

2. Explainable AI(XAI) XAI Inner Workings Prediction (Grad-CAM, Activation map, Cycle-GAN, etc ,.) Understanding Irr /Reducible Uncertainty (Finding outliers, aleatoric uncertainty.) Explaining Decisions Research papers published targeting XAI.(2006-2020), (Survey of XAI in digital pathology)

3. Explainable AI(XAI) L ocally- I nterpretable M odel-agnostic E xplanations (LIME) Perturbed Image Super- Pixels Square Image Grid of varying dimensions Local interpretable model-agnostic explanations for classication of lymph node metastases. Sensors 19 (13), 2969 ( jul 2019). https://doi.org/10.3390/s19132969 Measures which block/ superpixel (hence the cells) that has the most contribution in classifying a patch into a certain cancer type.

4. Explainable AI(XAI) Prediction and reconstruction Cycle-GAN Grad-CAM Cancerous Normal Cancerous Normal Brighter color represents where the NN is focusing in order to classify it as a cancerous region. Can explain the pixels or the area of pixels are most influencing the networks decision. CELNet : Evidence Localization for Pathology Images using Weakly Supervised Learning

5. Explainable AI(XAI) Uncertainty driven pooling network for microvessel segmentation in routine histology images Uncertainty Estimation model Uncertainty prediction Random augmentations Predicted Uncertainty Typical Uncertainty estimation process Helps us to explain the confidence in models prediction on pixel level.