Workshop on Interpretability of Machine Intelligence in Medical Image Computing at MICCAI 2020

Overview

Machine learning (ML) systems are achieving remarkable performances at the cost of increased complexity. Hence, they become less interpretable, which may cause distrust. As these systems are pervasively being introduced to critical domains, such as medical image computing and computer assisted intervention (MICCAI), it becomes imperative to develop methodologies to explain their predictions. Such methodologies would help physicians to decide whether they should follow/trust a prediction or not. Additionally, it could facilitate the deployment of such systems, from a legal perspective. Ultimately, interpretability is closely related with AI safety in healthcare.

However, there is very limited work regarding interpretability of ML systems among the MICCAI research. Besides increasing trust and acceptance by physicians, interpretability of ML systems can be helpful during method development. For instance, by inspecting if the model is learning aspects coherent with domain knowledge, or by studying failures. Also, it may help revealing biases in the training data, or identifying the most relevant data (e.g., specific MRI sequences in multi-sequence acquisitions). This is critical since the rise of chronic conditions has led to a continuous growth in usage of medical imaging, while at the same time reimbursements have been declining. Hence, improved productivity through the development of more efficient acquisition protocols is urgently needed.

The Workshop on Interpretability of Machine Intelligence in Medical Image Computing (iMIMIC) at MICCAI 2020 aims at introducing the challenges & opportunities related to the topic of interpretability of ML systems in the context of MICCAI.

Scope

Interpretability can be defined as an explanation of the machine learning system. It can be broadly defined as global, or local. The former explains the model and how it learned, while the latter is concerned with explaining individual predictions. Visualization is often useful for assisting the process of model interpretation. The model’s uncertainty may be seen as a proxy for interpreting it, by identifying difficult instances. Still, although we can find some approaches for tackling machine learning interpretability, there is a lack of formal and clear definition and taxonomy, as well as general approaches. Additionally, interpretability results often rely on comparing explanations with domain knowledge. Hence, there is the need for defining objective, quantitative, and systematic evaluation methodologies.

Covered topics include but are not limited to:

• Definition of interpretability in context of medical image analysis.
• Visualization techniques useful for model interpretation in medical image analysis.
• Local explanations for model interpretability in medical image analysis.
• Methods to improve transparency of machine learning models commonly used in medical image analysis.
• Textual explanations of model decisions in medical image analysis.
• Uncertainty quantification in context of model interpretability.
• Quantification and measurement of interpretability.
• Legal and regulatory aspects of model interpretability in medicine.

Program

The program of the workshop includes keynote presentations of experts working in the field of interpretability of machine learning. A selection of submitted manuscripts will be chosen for short oral presentations (10 minutes + 3 minutes Q&A) alongside the keynotes. Finally, we will have a group discussion which leaves room for a brainstorming on the most pressing issues in interpretability of machine intelligence in the context of MICCAI.

Final program:

UTC time - October 4th, 2020

Keynote speakers

• Himabindu Lakkaraju, Harvard University, USA.

  Title: Understanding the Limits of Explainability in ML-Assisted Decision Making

  

  Abstract: As machine learning black boxes are increasingly being deployed in domains such as healthcare and criminal justice, there is growing emphasis on building tools and techniques for explaining these black boxes in an interpretable manner. Such explanations are being leveraged by domain experts to diagnose systematic errors and underlying biases of black boxes. In this talk, I will demonstrate that post hoc explanations techniques that rely on input perturbations, such as LIME and SHAP, are not reliable. Specifically, I will discuss a novel scaffolding technique that effectively hides the biases of any given classifier by allowing an adversarial entity to craft an arbitrary desired explanation. Our approach can be used to scaffold any biased classifier in such a way that its predictions on the input data distribution still remain biased, but the post hoc explanations of the scaffolded classifier look innocuous. Using results from real world datasets (including COMPAS), I will demonstrate how extremely biased (racist) classifiers crafted by our framework can easily fool popular explanation techniques such as LIME and SHAP into generating innocuous explanations which do not reflect the underlying biases. I will conclude the talk by discussing extensive user studies that we carried out with domain experts in law to understand the perils of such misleading explanations and how they can be used to manipulate user trust.

• Wojciech Samek, Fraunhofer HHI, Germany.

  Title: Extending Explainable AI Beyond Deep Classifiers

  

  Abstract: Over the years, ML models have steadily grown in complexity, gaining predictivity often at the expense of interpretability. Explainable AI (or XAI) has emerged with the goal to produce models that are both predictive and understandable. XAI has reached important successes, such as robust heatmap-based explanations of DNN classifiers. From an application perspective, there is now a need to massively engage into new scenarios such as explaining unsupervised / reinforcement learning and non-neural network type models ML models, as well as producing explanations that are optimally structured for the human, and explanations. This talk will summarize recent development in extending XAI beyond deep classifiers.

Winner of a pecuniary award of 300€ for the best paper:

Projective Latent Interventions for Understanding and Fine-tuning Classifiers by Andreas Hinterreiter, Marc Streit, Bernhard Kainz.

Paper submission