Automated ROI Selection for Fatty Liver Disease Severity Classification Using Attention Map Analysis

Fatty liver disease (FLD) is a prevalent liver disease that often remains asymptomatic in the early stages. However, if left untreated and uncontrolled, it can progressively develop into a severe health problem. Clinical practice frequently relies on imaging and histological examinations. However, current approaches lack a cost-effective or non-invasive means of achieving highly accurate diagnoses. Ultrasonography offers a promising avenue in the field of imaging diagnostics. Nonetheless, previous studies focusing on the classification of fatty liver disease using ultrasonic images and machine learning have heavily relied on the manual delineation of regions of interest by expert physicians. This approach is susceptible to observer bias and is time-consuming and labor-intensive. To address this issue, this study proposes an algorithm for automatically searching for optimal region of interest (ROI). The objective is to automate ROI selection optimization, which can reduce the manual steps required in the model-building process and achieve effective discrimination of fatty liver severity. In this study, using only 205 cases, the proposed method achieved classification performance for mild and moderate fatty liver with F1-scores of 83.87% and 78.79%, and accuracies of 80.77% and 86.54%, respectively.