Florian Jung's research while affiliated with Fraunhofer Institute for Computer Graphics Research IGD and other places

... Various methods other than deep learning have been proposed for automatic segmentation of OAR in the head and neck region. The approaches include, among others, (multi) atlas-based methods, 3-6 model-based methods, [7][8][9][10] or their combinations. 11,12 Some of the methods 6,[8][9][10][11] have been evaluated in the 2015 MICCAI challenge on head and neck autosegmentation, where the best mean Dice score on the parotid glands was 0.84. ...

... There are several methods for processing the human teeth, including methods for teeth reconstruction [Abdelrehim et al. 2014;Farag et al. 2013;Wirtz et al. 2021;Wu et al. 2016;Zheng et al. 2011], restoration and completion [Mostafa et al. 2014;Ping et al. 2021], orthodontic treatment [Yang et al. 2020], segmentation Zhang et al. 2021], pose estimation [Beeler and Bradley 2014;Murugesan et al. 2018;Yang et al. 2019] and others [Velinov et al. 2018;Wei et al. 2020]. The closest to our work are methods for teeth reconstruction and restoration [Abdelrehim et al. 2014;Mostafa et al. 2014;Ping et al. 2021;Wirtz et al. 2021;Wu et al. 2016]. ...

... The suggested visual analytics interface can thus be used to extend approaches, which support retraining models with user-corrected image annotations such as the setup suggested by Dikici et al. (51). Related approaches for the application of visual analytics tools in the exploration of multimodal study data including image information as suggested by Bannach et al. (52) and Angulo et al. (53) strongly focus on the visualization of parameter distributions and have not been applied in a data curation context. However, our solution could also be used for cohort exploration and enhanced by more context-specific visualizations of the cardiac anatomy as suggested e.g., by Meuschke et al. (54). ...

... The field of damage recognition on built structures is still unexplored. In contrast to the fields of autonomous driving [12,18,9,34,43] or medicine [29,35,32], semantic segmentation benchmarks for damage recognition are rare. To the best of our knowledge, only two relevant benchmarks in the domain of reinforced concrete defects (RCDs) exist: CrackSeg9k [26] and S2DS [7]. ...

... The gross tumor volume was segmented at the clinical centers using a semi-automatic segmentation software based on coupled shape modeling [55]. The segmentation of the region of interest (ROI), corresponding to the primary tumor, was performed manually slice by slice by expert radiologists (one for each center) dedicated to head and neck cancers. ...

... We compare our proposed ECONet with existing state-of-the-art methods in online likelihood inference, which are Histogram (Boykov and Jolly, 2001), Gaussian Mixture Model (GMM) (Rother et al., 2004) and DybaORF-Haar-Like (Wang et al., 2016). In addition, to show the effectiveness of learning features in ECONet, we define ECONet-Haar-Like that replaces the first convolution layer of ECONet with hand-crafted haar-like features (Jung et al., 2013) and learns the three fully-connected layers. Both DybaORF-Haar-Like and ECONet-Haar-Like utilize our GPU-based implementation of 3d haar-like features, available at: https://github.com/masadcv/PyTorchHaarFeatures. ...

... Among the existing architectures, nn-UNet is commonly a top performer in prostate segmentation challenges and is considered as the de facto choice for the task [8,9]. Whilst a broad range of results have been presented for prostate WG segmentation trained on open-source and single-institution datasets [9,10], little is known about the impact of the MRI scanner characteristics and their degree of adherence to PI-RADS v2.1 technical standards on the inter-institutional transferability and longitudinal performance of the model after a successful deployment [11]. ...

... Manual contours of the prostate from MR examinations were used to assess each registration method. The aligned contours from each registration method were compared using the Dice similarity index (DSI) [27] for volume overlap (1 representing perfect overlap and 0 representing no overlap) and surface distance differences (mm) based on the 95% Hausdorff distance (HD) [28] and average surface distance (ASD) [29]. The HD is a measurement of the largest minimum distance between two contours. ...