Bachelor Thesis: Unsupervised 3D Clustering and Quantification of Pathological Hotspots in Orthopaedic SPECT/CT Imaging

12.05.2025, Diplomarbeiten, Bachelor- und Masterarbeiten

This project proposes an unsupervised AI-driven pipeline to detect and quantify pathological uptake hotspots in 3D orthopaedic SPECT/CT scans. It aims to enhance diagnostic objectivity by extracting detailed spatial and statistical features relevant to clinical decision-making.

Institute for AI and Informatics in Medicine

Chair for Orthopaedics and Sports Orthopaedics

Research group AI in Orthopaedics (AIO)

Title

Unsupervised 3D Clustering and Quantification of Pathological Hotspots in Orthopaedic SPECT/CT Imaging

Project Background

Single Photon Emission Computed Tomography combined with Computed Tomography (SPECT/CT) provides crucial functional and anatomical data for the assessment of orthopaedic pathologies. However, interpretation of these scans remains largely qualitative and observer-dependent. Recent work has proposed standardized, quantitative analysis pipelines, yet most approaches rely on predefined thresholds or regions of interest. This project seeks to develop an unsupervised learning pipeline capable of automatically identifying and quantifying abnormal uptake regions ("hotspots") in 3D SPECT/CT volumes. The model will extract clinically relevant statistics—such as volume, extent, intensity peak, swelling patterns, and signal variation (e.g., standard deviation)—to support more objective and reproducible diagnostic workflows.

The project will be conducted in close collaboration with the Department of Orthopaedics and Sports Orthopaedics at the University Hospital, the Institute for AI and Informatics in Medicine, and the Kantonsspital Baselland – Klinik für Orthopädie und Traumatologie des Bewegungsapparates. It will be supervised within the AI in Orthopaedics (AIO) group, an interdisciplinary team of clinicians and computer scientists affiliated with these institutions.

Your Tasks

• Design and implement an unsupervised 3D clustering algorithm (e.g., DBSCAN, Gaussian Mixture Models) to detect abnormal SPECT/CT hotspots.
• Develop a post-processing pipeline to quantify volumetric and statistical properties of the hotspots, including volume, swelling, standard deviation, and peak uptake.
• Validate the method against clinical annotations and evaluate robustness across anatomical regions and patient cohorts.
• Document findings in a reproducible format and prepare a scientific report or publication.

What We Offer

• Access to a curated dataset of clinical SPECT/CT scans with expert annotations.
• Guidance from both clinical and AI experts in a collaborative, interdisciplinary research setting.
• Opportunity to contribute to ongoing research and potential publication in medical imaging or orthopaedic AI journals.
• Exposure to cutting-edge AI tools and frameworks for 3D medical image analysis.

Project Details

The student will first explore the landscape of unsupervised clustering techniques in volumetric data and identify suitable algorithms for SPECT/CT intensity distributions. The core task involves implementing a pipeline that segments pathological hotspots without prior labels and calculates clinically relevant statistics. Evaluation will be done using available ground truth or expert-derived annotations, with attention to generalizability and interpretability of the results.

Short Summary

This project proposes an unsupervised AI-driven pipeline to detect and quantify pathological uptake hotspots in 3D orthopaedic SPECT/CT scans. It aims to enhance diagnostic objectivity by extracting detailed spatial and statistical features relevant to clinical decision-making.

References

1. Hirschmann, M. T., et al. (2012). Standardized volumetric 3D-analysis of SPECT/CT imaging in orthopaedics: overcoming the limitations of qualitative 2D analysis. BMC Medical Imaging, 12(5). https://doi.org/10.1186/1471-2342-12-5
2. Raza, K. and Singh, N.K., 2021. A tour of unsupervised deep learning for medical image analysis. Current Medical Imaging Reviews, 17(9), pp.1059-1077.
3. Zhao, B., Ren, Y., Yu, Z., Yu, J., Peng, T. and Zhang, X.Y., 2021. Aucseg: An automatically unsupervised clustering toolbox for 3d-segmentation of high-grade gliomas in multi-parametric mr images. Frontiers in Oncology, 11, p.679952.
4. Chen, J., Li, Y., Luna, L.P., Chung, H.W., Rowe, S.P., Du, Y., Solnes, L.B. and Frey, E.C., 2021. Learning fuzzy clustering for SPECT/CT segmentation via convolutional neural networks. Medical physics, 48(7), pp.3860-3877.

Contact

Contact: florian.hinterwimmer@tum.de

Application: Please include a short CV and a recent transcript of records with your inquiry.

Kontakt: florian.hinterwimmer@tum.de