Index

Evaluation of One-Stage Object Detectors for Implant Detection in Intraoperative CBCT Projection Images

Evaluation and Fusion of Vision-Language and Computer Vision Models for On-road Scenario Extraction in Autonomous Vehicles

Mutual Information-Based Segmentation for Unseen Domain Generalization in Digital Pathology

The introduction of automated slide scanners has facilitated the digitization of histopathological samples, enhancing the capabilities of traditional light microscopy by allowing the use of automated image analysis algorithms. Machine learning algorithms have demonstrated great potential in this regard by extrapolating learned characteristics from annotated datasets to unseen data, thus providing valuable assistance to pathologists in their diagnostic work. The performance of these models, however, can be significantly degraded by variations in image characteristics, including differences in scanners used for image acquisition, staining methods, resolution, illumination, and artifacts [1, 2]. These challenges highlight the difficulty of applying trained models across environments, necessitating domain adaption techniques.

Previous studies have already addressed color inconsistencies in histological samples, with calibration slides being one approach to resolving scanner-dependent variations [3]. Further notable pre-processing (-)/ training (⋆) techniques include:
– Data augmentation to simulate variability in the input data (e.g. domain-, spatial transformations) [4,5]
– Image-level domain adaption to align visual features across domains, mitigating distributional discrepancies, e.g. stain normalization to reduce inter-sample/ inter-scanner color variation [5,6]
– Multi-scale processing to capture features at different resolutions [2]
⋆ Heterogeneous dataset training to improve model generalization across multiple sources [7]
⋆ Transfer learning to utilize pre-trained models which is ideal for sparsely annotated data [2]
⋆ Domain-invariant feature learning to ensure robustness to scanner and staining variability [8,9], and in particular adversarial training to reinforce robustness against domain shifts [2]
⋆ Disentangled feature learning to isolate distinct underlying factors of data variations, compelling the network to learn shared statistical components across different domains [5]

This thesis investigates the applicability of a mutual information-based method for feature disentanglement [5] for cross-domain tumor segmentation in histopathology samples. By separating anatomical features from domain-specific variations, we aim for robust scanner-invariant segmentation performance. The objective is to enhance the generalizability of the network and enable direct application to unseen domains without adaptation.

The proposed work comprises the following work items:
– Literature review of device-induced variations in microscopy image data and state-of-the-art methods to address them
– Conceptualization and adaptation of mutual information-based segmentation [5] to address generalization for unseen domains in microscopy image data
– Exploration of targeted augmentation methods for addressing domain shifts in histopathology (e.g. stain augmentation [6])
– Exploration of suitable metrics for evaluating cross-domain generalization performance
– Documentation and presentation of the findings, documentation of code

[1] F. Wilm, M. Fragoso, C. A. Bertram, N. Stathonikos, M. Öttl, J. Qiu, R. Klopfleisch, A. Maier, K. Breininger, and M. Aubreville, “Multi-scanner canine cutaneous squamous cell carcinoma histopathology dataset,” in Bildverarbeitung für die Medizin 2023: Proceedings, German Workshop on Medical Image Computing, Braunschweig, July 2-4, 2023 (T. M. Deserno, H. Handels, A. Maier, K. Maier-Hein, C. Palm, and T. Tolxdorff, eds.), Informatik aktuell, Wiesbaden: Springer Fachmedien Wiesbaden, 2023.
[2] C. L. Srinidhi, O. Ciga, and A. L. Martel, “Deep neural network models for computational histopathology: A survey,” Medical Image Analysis, vol. 67, p. 101813, Jan. 2021.
[3] X. Ji, R. Salmon, N. Mulliqi, U. Khan, Y. Wang, A. Blilie, B. G. Pedersen, K. D. Sørensen, B. P. Ulhøi, R. Kjosavik, E. A. M. Janssen, M. Rantalainen, L. Egevad, P. Ruusuvuori, M. Eklund, and K. Kartasalo, “Physical Color Calibration of Digital Pathology Scanners for Robust Artificial Intelligence Assisted Cancer Diagnosis.”
[4] M. Balkenhol, N. Karssemeijer, G. J. S. Litjens, J. Van Der Laak, F. Ciompi, and D. Tellez, “H&E stain augmentation improves generalization of convolutional networks for histopathological mitosis detection,” in Medical Imaging 2018: Digital Pathology (M. N. Gurcan and J. E. Tomaszewski, eds.), (Houston, United States), p. 34, SPIE, Mar. 2018.
[5] Y. Bi, Z. Jiang, R. Clarenbach, R. Ghotbi, A. Karlas, and N. Navab, “MI-SegNet: Mutual Information-Based US Segmentation for Unseen Domain Generalization,” Feb. 2024. arXiv:2303.12649.
[6] M. Macenko, M. Niethammer, J. S. Marron, D. Borland, J. T. Woosley, Xiaojun Guan, C. Schmitt, and N. E. Thomas, “A method for normalizing histology slides for quantitative analysis,” in 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, (Boston, MA, USA), pp. 1107–1110, IEEE, June 2009.
[7] M. Aubreville, F. Wilm, N. Stathonikos, K. Breininger, T. A. Donovan, S. Jabari, M. Veta, J. Ganz, J. Ammeling, P. J. Van Diest, R. Klopfleisch, and C. A. Bertram, “A comprehensive multi-domain dataset for mitotic figure detection,” Scientific Data, vol. 10, p. 484, July 2023.
[8] A. Moyes, “A Novel Method For Unsupervised Scanner-Invariance With DCAE Model.”
[9] M. W. Lafarge, J. P. W. Pluim, K. A. J. Eppenhof, P. Moeskops, and M. Veta, “Domain-adversarial neural networks to address the appearance variability of histopathology images,” 2017. arXiv:1707.06183.

Agentic Radiology Report Generation

Large Language Models for Modified Frenchay Dysarthria Assessment Reports from Parkinson’s Speech: Model Choice and Prompting Effects

Advanced nnU-Net Ensemble Techniques for Lung Nodule Segmentation

This thesis outlines a comprehensive research program aimed at advancing medical
image segmentation through enhanced nnU-Net ensemble methodologies. Building upon
substantial experimental results that demonstrate significant improvement over existing
approaches, the proposed research addresses critical gaps in current medical imaging AI
capabilities.  The research aims to establish new performance standards for automated lung
nodule detection with current achievements of 0.84 dice coefficient representing a 29.2%
improvement over SAM baseline and approaching clinical utility thresholds. Future work will
focus on Vision Transformer integration with nnU-Net architectures, generalization validation
across additional lung imaging datasets, and clinical deployment optimization. The expected
outcomes include significant academic contributions through peer-reviewed publications,
practical clinical applications with potential for real-world healthcare impact, and establishment
of open-source implementations for research community adoption.

Development of a Local LLM Agent System for Clinical Expert Support and Automation in MRI Planning for Radiation Therapy

If you are interested, please contact fabian.wagner@fau.de

Cold Diffusion for CT Field-of-View Extension

Adaptive Biophysical Modelling for Thermal Ablation

Reinforcement Learning for Adaptive Protection in Power Grids

This thesis explores the use of reinforcement learning to improve protection strategies in power grids with high penetration of renewable energy. Conventional relay schemes often fail under changing fault conditions caused by inverter-based DERs. This thesis investigates how adaptive, data-driven control can overcome these challenges. A simulated environment based on DIgSILENT PowerFactory enables comparison between traditional protection and learning-based approaches.