NON-INVASIVE METHODS TO PREDICT AND MONITORING METABOLIC DYSFUNCTION ASSOCIATED STEATOTIC LIVER DISEASE BY USING ULTRASOUND IMAGING

Metabolic-Associated Steatotic Liver Disease (MASLD) is becoming increasingly critical due to its rising global prevalence and its potential to progress to more severe liver diseases, such as Metabolic-Associated Steatohepatitis (MASH), fibrosis, cirrhosis, and hepatocellular carcinoma (HCC). The early detection of MASLD is essential for a good management of patients and prevention of its progression. Traditional method such as liver biopsy is expensive, highly invasive and characterized by sampling errors. These issues highlight the need for non-invasive alternatives. Several medical imaging techniques are widely used in this setting. Accordingly, techniques such as ultrasound (US), magnetic resonance imaging (MRI), and transient elastography have shown promise in providing means for liver fat quantification and fibrosis assessment, which could significantly improve clinical outcomes for patients. Furthermore, advancements in image processing techniques, artificial intelligence (AI)-based (from machine learning to deep learning) models applied to imaging data give the potential to improve MASLD quantification, allowing for more widespread use of these technologies in both research and clinical settings and helping and supporting clinicians in their decision-making processes. With this regard, this PhD thesis investigates the use of conventional image processing techniques, machine learning (ML), and deep learning (DL) models to predict liver fat content (LFC) from US B-mode images, using magnetic resonance (MR) imaging LFC values as the reference standard. The primary aim that guided this PhD trip has been to propose non-invasive systems for hepatic steatosis evaluation, by developing predictive models that can reliably estimate LFC based on US images. After a comprehensive study of the state-of-the art methods for the quantification of LFC on US data, conventional image processing methods were employed to extract relevant features from the US images with some clinical application of the developed model. Then, the application of both ML and DL models have been also investigated to build predictive algorithms with similar aim of quantifying LFC using MRI fat% as reference standard. The thesis compares the performance of these models against MR-based fat quantification, identifying the strengths and limitations of each approach. The results demonstrate that all conventional processing approaches, ML and DL techniques show promise in accurately predicting LFC, offering a safe, fast, and low-cost alternative to traditional MR imaging and liver biopsy. The study concludes by discussing the potential for clinical implementation, challenges, and future directions in improving the accuracy and robustness of these models for widespread use in diagnostic practice, hoping at improving management of patients in routine clinical assessment and outcomes.