Development of AI-based Algorithm for the segmentation of biomedical Ultrasound images.

In this Thesis we propose an AI-based method to identify vertebrae within the lumbar region from ultrasound images, a fundamental step towards the diagno- sis of Osteoporosis. The measure of Bone Mineral Density (BMD) is an important operational index for the assessment of bone health status and osteoporosis in adults, which, within the innovative Radiofrequency Echographic Multi Spectrometry (REMS) diagnostic pipeline, is computed from the raw signals coming from the lumbar vertebrae. It is hence important to correctly identify the vertebrae with a segmentation step, which is currently based on traditional image segmentation methods, such as clustering and thresholding. Although these methods are robust, they have the characteristic of being highly dependent on the developer’s experience and skills in encoding expert knowledge into the segmentation algorithm. This limitation may impair the re-adaptation of the segmentation algorithm to other anatomic sites and/or other diagnostic populations such as infants and adolescents. The aim of the thesis is to develop an advanced AI-based data-driven segmentation approach capable of overcome this limitation while at the same time keep the robustness unaltered. The method developed is computationally efficient compared to classical segmentation approach. Indeed it eliminate the need for resource-intensive pre-processing and manual feature engineering, leading to substantial time and resource savings, especially in large-scale or real-time segmentation tasks. To further emphasize this efficiency, it’s important to note that the classical segmentator requires segmentation times on the order of tenth of a second per datapoint, while the CNN-based approach result to be one order of magnitude faster. In this context, we adopted the UNet architecture, which has been trained, validated, and tested against a population of 880 subjects. Once integrated into the REMS pipeline, it has been observed that the BMD values are comparable with the expected ones. Moreover, the new methodology is general enough to be easily transferred to other anatomic sites and/or other diagnostic populations.