Accuracy improvement of nuclear position extraction from hepatic histopathologic images

In the field of histopathologic diagnosis, differential diagnosis of borderline lesions is a serious problem. For example, differential diagnosis between early well-differentiated hepatocellular carcinoma (ewHCC) and noncancer is difficult because the cellular atypism of ewHCC is very low. Nuclear density (number of nuclei per unit area) is one of the most important features in differential diagnosis between non-cancer and ewHCC. We previously developed a support system for diagnosing ewHCC that enables the user to estimate the nuclear density. The system automatically extracts the positions of hepatocellular nuclei from a microscopic image. However, it takes a few minutes for the user to correct wrong positions using a graphical user interface (GUI). Our target is to improve the accuracy for nuclear position extraction and to make the system more convenient. To extract the positions of hepatocellular nuclei, a three-step method composed of nuclear candidates extraction, contours extraction and selection steps was used. In the nuclear candidates extraction step, lymphocytes, which are confusing with nuclei, were extracted and removed from the candidates to reduce wrongly extracted positions. In the contours extraction step, nucleoli, which sometimes lead to wrongly extracted contours, were identified. As a result, the accuracy for contours extraction was improved. Experimental results show the correction ratio (number of positions to be corrected / number of nuclei) was reduced to 19.1% from the conventional ratio of 31.3%. As a result, the time required to correct the wrong positions using a GUI was reduced to about 60% of conventional time. An experiment to differentially diagnose between ewHCC and non-cancer using the nuclear density was also carried out. 81% of true positive rate and 90% of true negative rate were obtained, and the usefulness of the nuclear density for the diagnosis was verified.