Collaborative slide screening for the diagnosis of breast cancer metastases in lymph nodes

Gianluca Gerard, Marco Piastra

Keywords: fully convolutional network, few-shot learning, meta-learning, sparse annotation, lymph nodes, camelyon16, histopathological images

TL;DR: Using collaborative fully convolutional networks for screening Whole Slide Images in histopathological diagnosis

Abstract:

In this paper we assess the viability of applying a few-shot algorithm to the segmentation of Whole Slide Images (WSI) for human histopathology. Our ultimate goal is to design a deep network that could screen large sets of WSIs of sentinel lymph-nodes by segmenting out areas with possible lesions. Such network should also be able to modify its behavior from a limited set of examples, so that a pathologist could tune its output to specific diagnostic pipelines and clinical practices.
In contrast, 'classical' supervised techniques have found limited applicability in this respect, since their output cannot be adapted unless through extensive retraining.
The novel approach to the task of segmenting biological images presented here is based on guided networks, which can segment a query image by integrating a support set of sparsely annotated images which can also be extended at run time.
In this work, we compare the segmentation performances obtained with guided networks to those obtained with a Fully Convolutional Network, based on fully supervised training. Comparative experiments were conducted on the public Camelyon16 dataset; our preliminary results are encouraging and show that the network architecture proposed is competitive for the task described.

Code Of Conduct: I have read and accept the code of conduct.