This master’s thesis by Jiale Hou, submitted to the Cyprus University of Technology in May 2025, addresses the challenge of automated prostate segmentation in ultrasound images. Prostate segmentation is a critical step in computer-aided diagnosis and treatment planning for prostate-related diseases, including cancer. Ultrasound imaging, due to its non-invasive nature and real-time capabilities, is widely used in clinical settings. However, the inherent noise, low contrast, and variability in ultrasound images make accurate segmentation a complex task. This thesis proposes enhanced methodologies that leverage diverse pre-processing strategies and multi-input neural network architectures to improve segmentation performance.
Diverse Pre-Processing Strategies: The thesis systematically investigates and implements multiple pre-processing techniques to address common issues in ultrasound imaging, such as speckle noise, intensity inhomogeneity, and boundary ambiguity. By optimizing these pre-processing steps, the quality of input data for segmentation models is significantly improved, leading to better delineation of the prostate region.
Multi-Input Architectures: Building on recent advances in deep learning, the work introduces and evaluates multi-input neural network architectures. These models are designed to process different representations or modalities of the input data simultaneously, enabling the network to learn complementary features and contextual information. This approach enhances the model’s ability to generalize across diverse ultrasound datasets and patient anatomies.
Comprehensive Evaluation: The thesis includes extensive experimental validation using clinical ultrasound datasets. Quantitative metrics such as Dice Similarity Coefficient, sensitivity, and specificity are reported, demonstrating the superiority of the proposed methods over conventional single-input and less sophisticated pre-processing approaches.
Clinical Relevance: By focusing on robust and automated solutions, the research aims to reduce inter-operator variability and improve the reproducibility of prostate segmentation in real-world clinical workflows.
The proposed enhancements in automated prostate segmentation have significant implications for both research and clinical practice. Improved segmentation accuracy facilitates more precise diagnosis, treatment planning, and monitoring of prostate diseases. The integration of advanced pre-processing and multi-input architectures sets a new benchmark for future studies in medical image analysis, particularly for challenging modalities like ultrasound. Furthermore, the methodologies developed in this thesis can be adapted to other organ segmentation tasks and imaging modalities, broadening their applicability. The work contributes to the ongoing efforts to harness artificial intelligence for improved healthcare outcomes, supporting the transition towards more personalized and data-driven medicine.