RADIOMICS ANALYSIS OF MRI LIVER FIBROSIS IN NASH PATIENTS

Abstract

Patients who suffer from non-alcoholic fatty liver disease (NON-ALCOHOLIC FATTY HEPATOPATHY) are the subject of extensive study at the moment. In this entity we differentiate those who only have fat, against those who develop non-alcoholic steatohepatitis (NASH), where fat produces lipotoxicity generating an inflammatory damage to the hepatocytes that condition an evolution to fibrosis of these patients and finally to cirrhosis and portal hypertension. The best resolution imaging method for the human body¿s interior is magnetic resonance imaging (MRI), however without first performing a fibroscan or biopsy, doctors are unable to visually assess the degree of fibrosis. Since biopsy has historically been the basis for diagnosis, there is interest in less invasive diagnostic methods that have hardly been created. There are currently open fronts in this conflict, but we want to use artificial intelligence and radiomics to attempt to offer a less intrusive alternative. Initially, a selection of images of two modalities, T2 and Diffusion (DWI), from a database of 31 patients diagnosed with NASH is performed to evaluate. They consist on 14 females and 20 males with an average age of 58.6 ± 11.9 years and a prevalence of obesity (grade II) by means of the Body Mass Index (BMI) average in the sample studied. A manual liver segmentation, extraction, and selection of radiomic features is carried out. Following the selection of these, the model will be evaluated by cross validation and training with a Random Forest before being evaluated through classification and regression. It will be considered the categorical objective value of the various degrees of fibrosis, ranging from 0 to 4 (4 being the greatest degree). Although it is not possible to compare a classification model with a regression model, due to their intrinsic differences, it can be seen how the metrics obtained in regression are quite fair in the case of Random Forest (regression), and very poor in the case of Linear Regression, given the non-linearity of the data. However, in the case of Random Forest classification, the performance of the model is quite high in both training and testing, with a sensitivity of 0.7 and specificity of 0.84, which indicates that it is the model that should ultimately be used to predict future data. In summary, the findings offer important new information about the application of AI and radiomics to the evaluation of liver fibrosis in NASH patients. It also provides avenues for further investigation, including the research of sophisticated feature selection methods and the integration of supplementary clinical data. In conclusion, this work lays a strong foundation for future investigations into the diagnosis and follow-up of patients with NASH.