Democritus University of Thrace (DUTH)
Χρήση αλγορίθμων τεχνητής νοημοσύνης για τη βελτιστοποίηση των εφαρμοζόμενων πρακτικών στην πυρηνική ιατρική
Abstract
dc:descriptionAim of the Study This thesis aims to develop and evaluate deep learning algorithms for denoising low-dose planar bone scintigraphy images. The main goal is to safely reduce the administered radiotracer activity and/or acquisition time while maintaining the diagnostic reliability of the images, thereby facilitating the practical implementation of the ALARA principle (As Low As Reasonably Achievable) and reducing radiation exposure for both patients and clinical staff. Methodology and Data The study was conducted using clinical data from 105 patients (44 males, 61 females; age range 37–88 years, mean age 67 ± 11.3) from two anatomical regions: thorax and pelvis. The images were acquired using Siemens gamma cameras with the radiopharmaceutical 99mTc-HDP at the Department of Nuclear Medicine of the University General Hospital of Alexandroupolis. Low-dose simulation was achieved by reducing acquisition time to 30%–90% of the standard dose. The dataset was divided into training (75%), validation (15%), and test (10%) subsets. Evaluation Procedure The evaluation consisted of two phases: 1.Comparative quantitative assessment of four algorithms: Four deep learning architectures (U-Net, DRUNet, MobileECADnCNN, and ECAE) were compared under identical conditions using PSNR and SSIM metrics. ECAE was identified as the best-performing algorithm. 2.In-depth evaluation of ECAE:oQuantitative analysis by anatomical region: PSNR and SSIM were computed for each dose level (30%–90%), revealing region-dependent performance differences. The pelvis, with more distinct anatomical structures, exhibited superior denoising compared to the thorax.oQualitative clinical assessment: Independent nuclear medicine physicians performed a qualitative evaluation using a Likert scale for four criteria (noise, anatomical visibility, detail preservation, and diagnostic suitability) and a Two-Alternative Forced-Choice (2AFC) test, demonstrating that the denoised images were visually preferred without loss of diagnostic information compared with full-dose images. Main ResultsThe ECAE outperformed the other algorithms both quantitatively and qualitatively, achieving the highest average PSNR (38.54 dB), compared to 37.79 dB for DRUNet and 37.63 dB for MobileECADnCNN, while maintaining comparable structural similarity (SSIM = 0.9513). Denoising significantly enhanced the visibility of anatomical structures and reduced noise, while preserving diagnostic information even at low-dose levels. Technical Features of ECAEECAE is an Enhanced Convolutional Autoencoder with multi-scale feature extraction and attention mechanisms, enabling the capture of long-range spatial dependencies between image regions and restoration of fine structural details in low-dose images. The incorporation of residual learning further improves the network’s learning capacity. Clinical Implications and ContributionThe study demonstrates that ECAE enables dose reduction without compromising diagnostic information, making the practical implementation of the ALARA principle feasible. Furthermore, the method can be extended to other nuclear medicine techniques (PET, SPECT, dynamic studies) and various radiopharmaceuticals, enhancing overall clinical safety and efficiency. Limitations and Future Directions The study was limited to planar images of the thorax and pelvis, without direct evaluation of specific pathological lesion detectability. Future research directions include extension to SPECT, additional anatomical regions, different radiopharmaceuticals, explainable artificial intelligence (AI) approaches, and task-based diagnostic assessment in real clinical settings. Conclusion This thesis provides a substantial contribution to the field of Nuclear Medicine by presenting a comprehensive methodology for dose reduction using deep learning, validated with real clinical data and a combination of quantitative and qualitative assessments. The work demonstrates the feasibility of safe and effective denoising, with significant benefits regarding optimization of patient safety and clinical protocols.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Bouzianis, Nikolaos
- Μπουζιάνης, Νικόλαος
Subjects
dc:subject × 22- Αλγόριθμοι βαθιάς μάθησης
- Αποθορυβοποίηση ιατρικής εικόνας
- Σπινθηρογράφημα οστών
- Πυρηνική ιατρική
- Ακτινοπροστασία
- Απεικόνιση μειωμένης δόσης
- Deep learning algorithms
- Medical image denoising
- Bone scintigraphy
- Nuclear medicine
- Radiation protection
- Low-dose imaging
- Επιστήμες Υγείας
- Ιατρική και Επιστήμες Υγείας
- Επαγγέλματα υγείας
- Κλινική Ιατρική
- Ραδιολογία, Πυρηνική ιατρική και Απεικονιστική
- Health Sciences
- Medical and Health Sciences
- Health Professions
- Clinical Medicine
- Radiology, Nuclear Medicine and Imaging
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61388