A Generic Neural Network Approach to Infer Segmenting Classifiers for Disease-Associated Regions in Medical Images

David Schuhmacher, Klaus Gerwert, Axel Mosig

Abstract:

In many settings in digital pathology or radiology, it is of predominant importance to train classifiers that can segment disease-associated regions in medical images. While numerous deep learning approaches, most notably U-Nets, exist to learn segmentations, these approaches typically require precise reference segmentations as training data. As a consequence, obtaining precise annotations of histopathological samples has become a major bottleneck to establish segmentation learning approaches. Our contribution introduces a neural network approach to avoid the annotation bottleneck in the first place: our approach requires two-class labels such as cancer vs. healthy at the sample level only. Using these sample-labels, a meta-network is trained that infers a segmenting neural network which will segment the disease-associated region (e.g. tumor) that is present in the cancer samples, but not in the healthy samples. This process results in a network, e.g. a U-Net, that can segment tumor regions in arbitrary further samples of the same type.

We establish and validate our approach in the context of digital label-free pathology, where hyperspectral infrared microscopy is used to segment and characterize the disease status of histopathological samples. Trained on a data set comprising infrared microscopic images of 100 tissue microarray spots labelled as either cancerous or cancer-free, the approach yields a U-Net that reliably identifies tumor regions or the absence of tumor in an independent test set involving 40 samples.

While our present work is focused on training a U-Net for infrared microscopic images, the approach is is generic in the sense that it can be adapted to other image modalities and essentially arbitrary segmenting network topologies.

TL;DR: Our approach allows to train U-Nets or arbitrary other segmenting classifiers without requiring ground truth segmentations, but using sample labels only.

Keywords: Image Segmentation, Deep Learning, Segmentation-free training