Emerging Topics in Deep Learning based Microscopy Image Analysis
MICCAI 2022 Tutorial
Scope
This tutorial will focus on emerging topics in deep learning-based microscopy image analysis. Recently, deep learning methodology for microscopy image analysis has evolved from application and optimisation of existing neural networks to more sophisticated models that are designed to target challenges in this problem domain more specifically and effectively. For instance, non-fully supervised learning approaches have recently attracted much attention to address the difficulty of limited training data; neural architecture search has been developed to enable automatic network optimisation; knowledge enhanced learning is another mechanism to enhance model performance without relying on excessive ground truth annotations; and many approaches have been developed to improve the robustness and interpretability of deep learning models. In this tutorial, we will present a broad and comprehensive overview of these various topics and engage researchers to discuss future research directions.
Learning Objectives
The tutorial aims to present a broad overview of several important emerging trends in deep learning for microscopy image analysis, providing a venue to summarise the current state-of-the-art as well as identifying promising future directions. Specifically, the tutorial will help the audience achieve the following goals:
Gain general knowledge of non-fully supervised learning in microscopy image analysis and gain familiarity with the main methods used in current state-of-the-art models;
Understand the main strategies of neural architecture search and analyse several recent studies in this domain;
Build a clear understanding of various types of knowledge enhanced learning and evaluate the advantages and disadvantages of different approaches for different problems; and,
Analyse issues of data imbalance, heterogeneity and interpretability and obtain an understanding of various techniques addressing these issues.
Organisers
Yang Song, University of New South Wales, Australia
Maurice Pagnucco, University of New South Wales, Australia
Erik Meijering, University of New South Wales, Australia
Weidong Cai, University of Sydney, Australia
Jianhua Yao, Tencent, China
Mei Chen, Microsoft, USA
Program Schedule
Sep 22, 2022 (Singapore time)
12:40-12:45pm: Opening [Slides]
12:45-1:15pm: Data-efficient learning (Yang Song) [Slides]
1:15-1:45pm: Neural architecture search (Erik Meijering) [Slides]
1:45-2:15pm: Knowledge representation and reasoning (Maurice Pagnucco) [Slides]
2:15-2:45pm: Microscopy imaging in industry (Jianhua Yao)
2:45-3:15pm: Robust and interpretable learning (Priyanka Rana, Cong Cong, Piumi Don Simonge)
3:15-3:20pm: Closing
Topic 1: Data-efficient learning
It is of great practical interest to develop deep learning methods that can work effectively without requiring extensive, precise manual annotations. We will introduce topics such as semi-supervised learning, weakly-supervised learning, self-supervised learning and unsupervised domain adaptation, and their applications in microscopy image analysis.
Topic 2: Neural architecture search (NAS)
Recently, NAS has been proposed as a mechanism to search for optimal network architectures automatically. We will give an overview of NAS and its application in microscopy image analysis, and discuss new developments in other domains that can inspire microscopy imaging research.
Topic 3: Knowledge enhanced learning
Incorporating knowledge priors can also help reduce the amount of training data required. We will discuss multiple ways of introducing knowledge, such as modelling contextual constraints and relationships, using domain ontologies to provide structured knowledge, and integrating learning with symbolic reasoning.
Topic 4: Robust and interpretable learning
Issues such as data imbalance and data heterogeneity are important problems to solve for enhancing method performance and robustness. Interpretability is often needed for model analysis and deployment. We will present recent studies and also discuss related methods in these areas.