Artificial Intelligence Techniques for Detecting Blood Vessels in Ultrasound Images

In recent years, medical science – and not only – has shifted towards a more personalized approach, relying on an individual’s profile rather than generalizations. This trend is also reflected in the present master's thesis, which leverages machine learning techniques to automate the detection of blood vessels in ultrasound images.

For this purpose, two datasets of medical ultrasound images were used. The first dataset focuses on the brachial plexus and includes 1,052 two-dimensional images from 101 patients, stored in JPEG format, with corresponding masks in JSON format. The second dataset focuses on the carotid artery and the internal jugular vein, containing 2,439 two-dimensional images from 15 participants, stored as NumPy arrays. In the first dataset, veins, arteries, muscles, and nerves were labeled (with this study focusing exclusively on veins and arteries), while the second dataset contained labels only for veins and arteries.

Initially, data preprocessing was performed, including the conversion of images to NPY format and the merging of the two datasets. Additional processing steps followed, which are described in detail in the chapters of this thesis. Subsequently, two deep learning models were selected for comparison: U-Net and SegNet, both of which are deep learning architectures commonly used for image segmentation. Their algorithms were developed and trained using the available data to achieve optimal blood vessel detection.

Finally, evaluation metrics were applied, highlighting not only the quality of the results but also their generalization capability. This further underscores the contribution of this study to the automation of medical diagnostics, making the process faster, easier, and more cost-effective.