Can weak segmentation labels from natural images aid digital pathology-based cancer classification?

Eu Wern Teh, Graham W. Taylor

Keywords: Transfer Learning, Digital Pathology

TL;DR: Aiding cancer classification in Digital Pathology with weak segmentation labels from natural image

Abstract: We explore the theme of utilizing weak segmentation labels from natural images to aid models perform better in Digital Pathology. A critical challenge with the medical image domain is the high cost of obtaining image labels from medical experts. On the other hand, it is considerably cheaper in time and money to obtain image labels from a layperson in the natural image domain. We show that general segmentation labels obtained from non-experts on a natural image dataset can boost performance on two cancer classification datasets (CRC and PCam). These results suggest that segmentation labels from natural images may not only benefit cancer classification but other future DP tasks.

Paper Type: validation/application paper

Primary Subject Area: Transfer Learning and Domain Adaptation

Secondary Subject Area: Application: Histopathology

Paper Status: original work, not submitted yet

Source Code Url: https://github.com/euwern/hist_weak_seg

Data Set Url: CRC dataset: https://zenodo.org/record/53169, PCam dataset: https://github.com/basveeling/pcam, Pascal VOC 2012 dataset: http://host.robots.ox.ac.uk/pascal/VOC/voc2012/

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.