NHS England Data Science PhD Internships

Explaining facial skin disease classification using LIME

Keywords: Explainability, LIME, Images

Need: Skin diseases are currently ranked as the fourth leading cause of non-fatal diseases worldwide with a significant treatment burden. Skin disease classification has been previously suggested as a comparable screening mechanism for skin cancers to current manual practice. Deep learning algorithms often appear as black-box and lack the explainability required to use them in a clinical setting. This project would seek to investigate applying a classification algorithm to skin-disease images and then demonstrating the use of LIME as an explainability tool.

See outputs from the project - LIME-XAI-Facial-Disease-Classification

Current Knowledge/Examples & Possible Techniques/Approaches: In May 2021, Google announced an AI-powered dermatology tool (paper here) built on a knowledge base of 288 skin conditions which can analyse a variety of skin conditions. The system is developed as a web-based application which enables a camera to take an image and then provides possible matching conditions of the diseases with further information.

To enable confidence and explainability in a deep learning model various model-agnostic methods can be investigated. Local Interpretable Model-agnostic Explanations (LIME) can be used to explain individual predictions and is applicable to high-dimensional data such as images.

There are a few works on LIME used for skin cancer image explanation. For example, Stieler et al. used LIME as an explainer for skin image classifiers. Xiang et al. used LIME for extracting evidence from skin images to support classification results.

Related Previous Internship Projects: n/a

Enables Future Work: Demonstration of applying model-agnostic explainability in a healthcare context.

Outcome/Learning Objectives: Use of Google Inception v3 as a base model for classifying facial skin conditions and then apply LIME as a surrogate model for explaining the predicted output.

Datasets: Open data such as http://www.dermweb.com/photo_atlas/; https://nsufl.libguides.com/c.php?g=112152&p=724706, or MIMIC-CXR.

Desired skill set: When applying please highlight any experience around image recognition, use of model-agnostic explainability, coding experience (including any coding in the open), any other data science experience you feel relevant.

Return to list of all available projects.