Keywords: whole-slide-image, domain adaptation, classification, deep learning.
TL;DR: A domain adaptation solution to WSI classfication
Abstract: Image classification on whole-slide-image (WSI) is a challenging task. A previous work based on Fisher vector encoding provided a novel end-to-end pipeline with promising accuracy and computational efficiency. However, this pipeline suffers from an accuracy drop when deployed to another dataset to perform the same task. This poses a limitation on the practical use of the pipeline especially when the diagnoses of WSIs are hard to obtain. This paper aims at providing a solution to mitigate the accuracy drop by using an unsupervised domain adaptation approach. We propose to insert the domain classifiers into the pipeline in two stages to align the features during training. We evaluate accuracy by calculating the confusion matrices before and after the adaptation on two datasets. We demonstrate that placing domain classifiers in different stages will boost accuracy.
Paper Type: methodological development
Source Latex: zip
Primary Subject Area: Transfer Learning and Domain Adaptation
Secondary Subject Area: Detection and Diagnosis
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/yuchen2580/double_adaptation_WSI
Data Set Url: For the target dataset, we use the data from warwick contest: https://warwick.ac.uk/fac/cross_fac/tia/data/her2contest/ The source dataset is from Alberta CCI which is not released to the public at this moment and awaiting approvals.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.