BEHSOF: Advanced Non-alcoholic fatty liver dataset with clinical metadata and ultrasound images for Deep learning Models

The liver is regarded as one of the essential organs in the body, playing a crucial role in digestion, nutrient absorption, and food processing. This indispensable organ is tasked with cleansing the blood that flows from the digestive tract. Additionally, the liver detoxifies harmful substances and metabolizes various medications. A key function of this organ involves intricate metabolic processes that transform food into energy. However, the liver's heightened sensitivity makes it susceptible to a variety of common ailments, underscoring the need for careful attention to its health. Among liver diseases, fatty liver stands out as the most prevalent, characterized by the buildup of fat within liver cells. This condition is particularly common among individuals who are overweight or have abdominal obesity. Specifically, non-alcoholic fatty liver disease (NAFLD) refers to the excessive fat accumulation in liver cells, a condition known as steatosis. Various diagnostic techniques have been developed for this disease, each offering distinct advantages and drawbacks. Ultrasound imaging has gained popularity due to its accessibility, non-invasive nature, and affordability. One of the approaches utilized in this diagnostic process is the employment of deep learning models. Sporadic studies have been conducted to leverage the available data in diagnosing the degree of non-alcoholic fatty liver disease, and given the significance of the issue, this body of research is continually expanding. Developing robust model training and conducting comprehensive analyses necessitates access to a diverse and comprehensive data repository that can be leveraged for training and validating the proposed models. To this end, we present the BEHSOF dataset, which comprises a collection of samples gathered with varying degrees of non-alcoholic fatty liver disease. Specifically, this data bank consists of ultrasound images from a population of 113 individuals under study, along with the corresponding labels for the levels of steatosis and fibrosis. In addition to the ultrasound images, the data bank provides clinical information, blood test results, and Fibroscan outcomes for the participants, which serve as reference data for fibrosis assessment. Finally, we showcase the results of two deep learning models as examples for training and testing the introduced data set.

肝脏被视作人体不可或缺的核心器官之一，在消化、营养吸收与食物代谢过程中发挥关键作用。这一不可或缺的器官肩负着净化来自消化道血液的重任，此外，肝脏还可对有害物质进行解毒，并代谢各类药物。该器官的核心功能之一，是通过复杂的代谢过程将食物转化为能量。然而，肝脏的高敏感性使其极易罹患多种常见疾病，这凸显了关注肝脏健康的必要性。在各类肝脏疾病中，脂肪肝最为高发，其病理特征为肝细胞内脂肪堆积。该病症在超重或腹部肥胖人群中尤为多见。具体而言，非酒精性脂肪肝（non-alcoholic fatty liver disease, NAFLD）指的是肝细胞内出现过量脂肪堆积的情况，即脂肪变性（steatosis）。目前已开发出多种针对该疾病的诊断技术，各有其独特的优势与局限。超声成像因操作便捷、无创且成本低廉而得到广泛应用。在此类诊断流程中，深度学习模型是常用的技术手段之一。已有少量研究利用现有数据开展非酒精性脂肪肝严重程度的诊断工作，鉴于该问题的重要性，相关研究正持续推进与拓展。构建可靠的模型训练体系并开展全面分析，需要一个多样化且全面的数据集仓库，用于训练与验证所提出的模型。为此，我们推出了BEHSOF数据集，该数据集涵盖了不同严重程度非酒精性脂肪肝的样本集合。具体而言，该数据库包含113名研究受试者的超声图像，以及对应的脂肪变性与纤维化程度标签。除超声图像外，该数据库还提供了受试者的临床信息、血液检验结果以及Fibroscan检测结果，作为纤维化评估的参考数据。最后，我们展示了两个深度学习模型的训练与测试结果，作为本数据集的应用示例。