back to list

Project: Visual Analytics for comparison of mitosis detection models

Description

Recently, A group of researchers including Mitko Veta (BMT) has organized the Mitosis Domain Generalization (MIDOG) challenge (more information here and here). The goal of the challenge was to assess the performance of algorithms for mitosis detection in histopathology images on both seen (available at training time) and unseen domains (not available at training time). As the appearance of histopathology images is greatly affected by the type of scanner used to produce the digital histopathology slides, we defined domains as images produced by one scanner (i.e. each scanner constitutes a domain). The challenge was very successful -  14 teams submitted methods for evaluation, with the top teams achieving accuracy comparable to that of experth pathologists on boths unseen and seen domains. 

Despite this success it remains difficult to identify what are the main properties that make an approach performe better and what are the  strengths and weaknesses of the different approaches. 

The goal of this project is to create visual analytics methods that helps identify the “strenghts and weakensess” of submitted methods with respect to the different test domains, as well as the apperance of mitotic figures, which is very heterogeneous. These methods can be used to correlate the performance of the method to specific design choices (e.g. neural network architecture, domain adaptation strategy, domain augmentation etc.) and inform development of methods with improved performance.

Visual Analytics has proven useful for model comparison when the goal is to go beyond quantification of performance [1][2]. However, the existing solutions are not sufficient to provide insight in the presented problem.  They do  not consider the specific complexity of the histopathology slides,  the domain generalizaiton and do not scale to the comparison beyond two models. 

References

[1] SUN D., FENG Z., CHEN Y., WANG Y., ZENG J., YUAN M., PONG T.-C., QU H.: DFSeer: A Visual Analytics Approach to Facilitate Model Selection for Demand Forecasting. Association for Computing Machinery, New York, NY, USA, 2020, p. 1–13. URL: https://doi.org/10.1145/3313831.3376866. 2

[2] ZHANG J., WANG Y., MOLINO P., LI L., EBERT D. S.: Manifold: A model-agnostic framework for interpretation and diagnosis of machine learning models. IEEE Transactions on Visualization and Computer Graphics 25, 1 (2019), 364–373. doi:10.1109/TVCG. 2018.2864499. 2

Details

Supervisor

Anna Vilanova

Secondary supervisor

MV

Mitko Veta (BMT)

Interested?

Get in contact