Understanding spurious decisions of Deep Neural Networks for Digital Dermatology

AI Quick Summary

This project aims at improving our understanding of how Deep Neural Networks make decisions on dermatologic datasets, which is critical for deployment. More precisely, the goal is to analyse different spurious signals such as hair, black corners, or ink marks that a Machine Learning model may use to infer if a mole is malignant in dermoscopic images.Results will support a collaboration between HSLU and the University Hospital of Basel for the development of new tools in Digital Dermatology. IntroductionMachine Learning applications to Digital Dermatology can disrupt the field by allowing fast, scalable solutions to improve healthcare services. This includes providing underdeveloped countries with telemedicine services.However, the allegedly most accurate method we have for the analysis of dermatologic images, Deep Neural Networks, cannot be directly interpreted by humans. Even more disturbingly, Deep Neural Networks seem to exploit patterns in data which have nothing to do with the criteria used by clinicians.GoalThis project aims at improving our understanding of how Deep Neural Networks make decisions on dermatologic datasets, which is critical for deployment. More precisely, the goal is to analyse different spurious signals such as hair, black corners, or ink marks that a Machine Learning model may use to infer if a mole is malignant in dermoscopic images.MethodsThe analysis of potential spurious biases in the ISIC dataset which presented in the two papers:Bissoto et al, (De)Constructing Bias on Skin Lesion Datasets, 2019Bissoto et al, Debiasing Skin Lesion Datasets and Models? Not So Fast, 2020will be partly reproduced and extended with new baselines.Moreover, evaluation will be repeated with models pre-trained using self-supervised learning on dermatologic datasets, including Convolutional Neural Networks and Vision Transformers.CollaborationResults will support a collaboration between HSLU and the University Hospital of Basel for the development of new tools in Digital Dermatology.

Project Idea Metadata

Project Idea Name: Understanding spurious decisions of Deep Neural Networks for Digital Dermatology
Date: 7/11/2022 3:27:09 PM
Administrators:

Project Idea Description

Introduction

Machine Learning applications to Digital Dermatology can disrupt the field by allowing fast, scalable solutions to improve healthcare services. This includes providing underdeveloped countries with telemedicine services.

However, the allegedly most accurate method we have for the analysis of dermatologic images, Deep Neural Networks, cannot be directly interpreted by humans. Even more disturbingly, Deep Neural Networks seem to exploit patterns in data which have nothing to do with the criteria used by clinicians.

Goal

Methods

The analysis of potential spurious biases in the ISIC dataset which presented in the two papers:

Bissoto et al, (De)Constructing Bias on Skin Lesion Datasets, 2019
Bissoto et al, Debiasing Skin Lesion Datasets and Models? Not So Fast, 2020

will be partly reproduced and extended with new baselines.

Moreover, evaluation will be repeated with models pre-trained using self-supervised learning on dermatologic datasets, including Convolutional Neural Networks and Vision Transformers.

Collaboration

Results will support a collaboration between HSLU and the University Hospital of Basel for the development of new tools in Digital Dermatology.