Authors:
Alessa Stria
and
Asan Agibetov
Affiliation:
Medical University of Vienna, Center for Medical Statistics, Informatics and Intelligent Systems (CeMSIIS), Institute of Artificial Intelligence, Vienna, Austria
Keyword(s):
Deep Learning, Segmentation, Classification, Explainable AI, Class Activation Map, Labeling Costs, Scarce Data, Sample Size Dependence, MRI, Cardiology.
Abstract:
Provided with a sufficient amount of annotated data, deep learning models have been successfully applied to automatically segment cardiac multi-structures from MR images. However, manual delineation of cardiac anatomical structures is expensive to acquire and requires expert knowledge. Recently, weakly- and self-supervised feature learning techniques have been pro-posed to avoid or substantially reduce the effort of manual annotation. Due to their end-to-end design, many of these techniques are hard to train. In this paper, we propose a simple modular segmentation framework based on U-net architecture that injects class activation maps of separately trained classification models to guide the segmentation process. In a small data setting (20-35% of training data), our framework significantly improved the segmentation accuracy of a baseline U-net model (5%-150%).