Keywords: guided network, conditional network, fully convolutional network, few-shot learning, meta-learning, sparse annotation, lymph nodes, camelyon16, histopathological images
TL;DR: Evaluation of the conditional FCN architecture performances in screening Whole Slide Images of lymphnodes based on a limited set of guiding examples.
Abstract: We assess the viability of applying a few-shot algorithm to the segmentation of Whole Slide Images (WSI) for human histopathology. The specific field considered is finding metastatic lesions in sentinel lymph-nodes. Given the huge size of WSIs and the substantial effort required by human pathologists to analyze them for diagnostic purposes, the goal is to design a system that could perform an automatic screening by segmenting out those areas that contain elements of potential interest. 'Classical' supervised techniques have found limited applicability in this respect, since their output cannot be adapted unless through extensive retraining. The approach to segmentation of histopathological images presented here is based on conditional FCN (co-FCN) networks, in which a fully convolutional network conditioned on an annotated support set of images do inference on an unannotated query image. After a complete end-to-end training, it is possible to correct the behavior at run time of a co-FCN by extending the support set, without further optimization. The adoption of co-FCN is expected to ease the annotation task and also to improve the acceptance by human experts, who will be able to correct the co-FCN behavior incrementally. In this preliminary work we use the publicly-available Camelyon16 dataset to show that the segmentation produced by co-FCN trained using late fusion and sparse annotations can be effectively modified at runtime, by integrating corrections on the fly.
Code Of Conduct: I have read and accept the code of conduct.
Remove If Rejected: Remove submission from public view if paper is rejected.