Nuclei detection and segmentation in hematoxylin and eosin-stained (H&E) tissue images are important clinical tasks and crucial for a wide range of applications. However, it is a challenging task due to nuclei variances in staining and size, overlapping boundaries, and nuclei clustering. While convolutional neural networks have been extensively used for this task, we explore the potential of Transformer-based networks in combination with large scale pre-training in this domain. Therefore, we introduce a new method for automated instance segmentation of cell nuclei in digitized tissue samples using a deep learning architecture based on Vision Transformer called CellViT. CellViT is trained and evaluated on the PanNuke dataset, which is one of the most challenging nuclei instance segmentation datasets, consisting of nearly 200,000 annotated nuclei into 5 clinically important classes in 19 tissue types. We demonstrate the superiority of large-scale in-domain and out-of-domain pre-trained Vision Transformers by leveraging the recently published Segment Anything Model and a ViT-encoder pre-trained on 104 million histological image patches - achieving state-of-the-art nuclei detection and instance segmentation performance on the PanNuke dataset with a mean panoptic quality of 0.50 and an F₁-detection score of 0.83. The code is publicly available at https://github.com/TIO-IKIM/CellViT.

Data mining of highly structured radiological reports from CT angiography enabled the development of a semiquantitative scoring system for thrombus load in pulmonary embolism. Background Many studies emphasize the role of structured reports (SRs) because they are readily accessible for further automated analyses. However, using SR data obtained in clinical routine for re...

Detector-based spectral CT offers the possibility of obtaining spectral information from which discrete acquisitions at different energy levels can be derived, yielding so-called virtual monoenergetic images (VMI). In this study, we aimed to develop a jointly optimized deep-learning framework based on dual-energy CT pulmonary angiography (DE-CTPA) data to generate synthetic monoenergetic images (SMI) for improving automatic pulmonary embolism (PE) detection in single-energy CTPA scans. For this purpose, we used two datasets: our institutional DE-CTPA dataset $D_1$, comprising polyenergetic arterial series and the corresponding VMI at low-energy levels (40 keV) with 7892 image pairs, and a 10% subset of the 2020 RSNA Pulmonary Embolism CT Dataset $D_2$, which consisted of 161,253 polyenergetic images with dichotomous slice-wise annotations (PE/no PE). We trained a fully convolutional encoder-decoder on $D_1$ to generate SMI from single-energy CTPA scans of $D_2$, which were then fed into a ResNet50 network for training of the downstream PE classification task. The quantitative results on the reconstruction ability of our framework revealed high-quality visual SMI predictions with reconstruction results of 0.984 ± 0.002 (structural similarity) and 41.706 ± 0.547 dB (peak signal-to-noise ratio). PE classification resulted in an AUC of 0.84 for our model, which achieved improved performance compared to other naïve approaches with AUCs up to 0.81. Our study stresses the role of using joint optimization strategies for deep-learning algorithms to improve automatic PE detection. The proposed pipeline may prove to be beneficial for computer-aided detection systems and could help rescue CTPA studies with suboptimal opacification of the pulmonary arteries from single-energy CT scanners.

Purpose: Interpreting chest radiographs (CXR) remains challenging due to the ambiguity of overlapping structures such as the lungs, heart, and bones. To address this issue, we propose a novel method for extracting fine-grained anatom- ical structures in CXR using pseudo-labeling of three-dimensional computed tomography (CT) scans. Methods: We created a large-scale dataset of 10,021 thoracic CTs with 157 labels and applied an ensemble of 3D anatomy segmentation models to extract anatomical pseudo-labels. These labels were projected onto a two-dimensional plane, similar to the CXR, allowing the training of detailed semantic segmentation models for CXR without any manual annotation effort. Results: Our resulting segmentation models demonstrated remarkable performance on CXR, with a high average model-annotator agreement between two radiologists with mIoU scores of 0.93 and 0.85 for frontal and lateral anatomy, while inter-annotator agreement remained at 0.95 and 0.83 mIoU. Our anatom- ical segmentations allowed for the accurate extraction of relevant explainable medical features such as the cardio-thoracic-ratio. Conclusion: Our method of volumetric pseudo-labeling paired with CT projection offers a promising approach for detailed anatomical segmentation of CXR with a high agreement with human annotators. This technique may have important clinical implications, particularly in the analysis of various thoracic pathologies.

Background: Pressure ulcers (PUs) and incontinence-associated dermatitis (IAD) are prevalent conditions in clinical settings, posing significant challenges due to their similar presentations but differing treatment needs. Accurate differentiation between PUs and IAD is essential for appropriate patient care, yet it remains a burden for nursing staff and wound care experts. Objective: This study aims to develop and introduce a robust multimodal deep learning framework for the classification of PUs and IAD, along with the fine-grained categorization of their respective wound severities, to enhance diagnostic accuracy and support clinical decision-making. Methods: We collected and annotated a dataset of 1555 wound images, achieving consensus among 4 wound experts. Our framework integrates wound images with categorical patient data to improve classification performance. We evaluated 4 models—2 convolutional neural networks and 2 transformer-based architectures—each with approximately 25 million parameters. Various data preprocessing strategies, augmentation techniques, training methods (including multimodal data integration, synthetic data generation, and sampling), and postprocessing approaches (including ensembling and test-time augmentation) were systematically tested to optimize model performance. Results: The transformer-based TinyViT model achieved the highest performance in binary classification of PU and IAD, with an F1-score (harmonic mean of precision and recall) of 93.23%, outperforming wound care experts and nursing staff on the test dataset. In fine-grained classification of wound categories, the TinyViT model also performed best for PU categories with an F1-score of 75.43%, while ConvNeXtV2 showed superior performance in IAD category classification with an F1-score of 53.20%. Incorporating multimodal data improved performance in binary classification but had less impact on fine-grained categorization. Augmentation strategies and training techniques significantly influenced model performance, with ensembling enhancing accuracy across all tasks. Conclusions: Our multimodal deep learning framework effectively differentiates between PUs and IAD, achieving high accuracy and outperforming human wound care experts. By integrating wound images with categorical patient data, the model enhances diagnostic precision, offering a valuable decision-support tool for health care professionals. This advancement has the potential to reduce diagnostic uncertainty, optimize treatment pathways, and alleviate the burden on medical staff, leading to faster interventions and improved patient outcomes. The framework’s strong performance suggests practical applications in clinical settings, such as integration into hospital electronic health record systems or mobile applications for bedside diagnostics. Future work should focus on validating real-world implementation, expanding dataset diversity, and refining fine-grained classification capabilities to further enhance clinical utility.