Identification of a suitable transfer learning architecture for classification: A case study with liver tumors

Recently liver diseases are emerging as a predominant medico-social problem in places where there is high prevalence of alcoholism and among the sub-set of people with unhealthy dietary habits. The incidence and mortality rates of liver cancer witness an increasing trend year by year globally and have almost tripled in the past forty years. As no symptoms are exhibited in early stages, liver cancer is diagnosed only at a later stage. Computed tomography (CT) is used as the primary imaging modality for the diagnosis of hepatocellular carcinoma (HCC), a primary liver cancer and other liver related disease. The CT image as such does not provide any clinical information pertaining to liver cancer/tumor, and hence, an intravenous iodinated contrast agent is injected prior to CT acquisition for the purpose of highlighting the tumorous tissue from the healthy liver. Accordingly, contrast enhanced computed tomography (CECT) images are acquired to make the tumorous tissue to be predominantly visible and influence well during the clinical diagnosis. In spite of the herculean visualization, CECT at times fails to provide clear picture of the abnormal parts of the liver which is otherwise called as lesions, making the diagnosis sub-optimal.