Meta-Learners for Few-Shot Weakly-Supervised Medical Image Segmentation

• DOI: 10.48550/arxiv.2305.06912
• Publication Date: 2023-01-01
• Authors: Hugo Oliveira; Pedro H. T. Gama; Isabelle Bloch; Roberto M. César
• Keywords: Modalities

Most uses of Meta-Learning in visual recognition are very often applied to image classification, with a relative lack works other tasks {such} as segmentation and detection. We propose generic framework for few-shot weakly-supervised medical imaging domains. conduct comparative analysis meta-learners from distinct paradigms adapted different sparsely annotated radiological tasks. The modalities include 2D chest, mammographic dental X-rays, well slices volumetric tomography resonance images. Our experiments consider total 9 meta-learners, 4 backbones multiple target organ explore small-data scenarios radiology varying weak annotation styles densities. shows that metric-based meta-learning approaches achieve better results smaller domain shifts comparison the meta-training datasets, while some gradient- fusion-based more generalizable larger shifts.