Mask Mining for Improved Liver Lesion Segmentation

Karsten Roth,Jürgen Hesser,Tomasz Konopczyński
DOI: https://doi.org/10.48550/arXiv.1908.05062
2020-03-12
Abstract:We propose a novel procedure to improve liver and lesion segmentation from CT scans for U-Net based models. Our method extends standard segmentation pipelines to focus on higher target recall or reduction of noisy false-positive predictions, boosting overall segmentation performance. To achieve this, we include segmentation errors into a new learning process appended to the main training setup, allowing the model to find features which explain away previous errors. We evaluate this on semantically distinct architectures: cascaded two- and three-dimensional as well as combined learning setups for multitask segmentation. Liver and lesion segmentation data are provided by the Liver Tumor Segmentation challenge (LiTS), with an increase in dice score of up to 2 points.
Image and Video Processing,Computer Vision and Pattern Recognition
What problem does this paper attempt to address?