In this work, the Menze group proposes a novel Detection Transformer for 3D anatomical structure detection, dubbed Focused Decoder. Focused Decoder precisely focuses on relevant anatomical structures, using anatomical region atlas information to deploy query anchors, while at the same time restricting the cross-attention’s field of view to regions of interest. Evaluated on two publicly available CT datasets, Focused Decoder is shown to provide both strong detection results, as well as highly intuitive explainability via attention weights.
See Wittmann et al.,Journal of Machine Learning for Biomedical Imaging
Code is available at https://github.com/bwittmann/transoar