3D characterization of thyroid tumors using micro-CT and machine learning
DOKPE
- Tajbakhsh, Kiarash ORCID University of Fribourg
- Fribourg (Switzerland) : Université de Fribourg, December 2025
1 ressource en ligne (xviii, 139 pages) ; 1 fichier pdf
PhD: Université de Fribourg (Suisse), 16.12.2025
3D virtual histology
Micro-CT
Propagation based phase-contrast imaging
Follicular thyroid neoplasms
Radiomics
Deep learning
English
This thesis develops advanced methodologies for the classification, diagnosis, and prognosis of thyroid tumors by integrating high-resolution X-ray imaging with machine learning. The first part focuses on optimizing micro-CT imaging of FFPE tissue blocks to improve contrast, spatial resolution, and reduce imaging artifacts. This enables 3D virtual histology and addresses limitations of conventional histology arising from sampling bias, particularly in follicular thyroid neoplasms. By examining entire tumor volumes, the approach enhances detection of capsular and vascular invasion. The methodology is further applied to clinically misclassified relapse cases to assess its potential for correcting diagnostic errors.
A comprehensive dataset of follicular thyroid neoplasms with expert annotations of capsular and vascular invasion is established. This dataset is used to train deep learning models for automated segmentation of invasive tumor features. The resulting framework supports pathologists by guiding targeted histological sectioning, improving diagnostic efficiency and reducing workload in clinical practice.
The second part extends the analysis to thyroid tumors of follicular origin, investigating surrogate imaging biomarkers for the prediction of clinically relevant genetic mutations with prognostic and therapeutic significance. Tumor subtyping is performed using texture-based analysis to capture intratumoral heterogeneity. Next-generation tissue microarrays enable efficient acquisition of diverse tumor samples. Radiomics and machine learning techniques are applied to extract quantitative features and build predictive models for tumor classification and molecular profiling. Overall, this work presents a translational framework combining advanced X-ray imaging and artificial intelligence for more accurate and personalized thyroid tumor management.
A comprehensive dataset of follicular thyroid neoplasms with expert annotations of capsular and vascular invasion is established. This dataset is used to train deep learning models for automated segmentation of invasive tumor features. The resulting framework supports pathologists by guiding targeted histological sectioning, improving diagnostic efficiency and reducing workload in clinical practice.
The second part extends the analysis to thyroid tumors of follicular origin, investigating surrogate imaging biomarkers for the prediction of clinically relevant genetic mutations with prognostic and therapeutic significance. Tumor subtyping is performed using texture-based analysis to capture intratumoral heterogeneity. Next-generation tissue microarrays enable efficient acquisition of diverse tumor samples. Radiomics and machine learning techniques are applied to extract quantitative features and build predictive models for tumor classification and molecular profiling. Overall, this work presents a translational framework combining advanced X-ray imaging and artificial intelligence for more accurate and personalized thyroid tumor management.
- Faculty
- Faculté des sciences et de médecine
- Department
- Département de Mathématiques
- Language
-
- English
- Classification
- Mathematics
- Notes
-
- Bibliographie
- License
-
CC BY
- Open access status
- diamond
- Identifiers
-
- SWISSCOVERY 991172740437005501
- DOI 10.51363/unifr.sth.2026.067
- URN urn:nbn:ch:rero-002-123536
- Persistent URL
- https://folia.unifr.ch/unifr/documents/334182
Statistics
Document views: 39
File downloads:
- phd_Tajbakhshk.pdf: 103