An Efficient Computer-Aided Diagnosis System for the Detection of Hepatocellular Carcinoma Nuclei Using a Convolutional Neural Network with Handcrafted Feature Injection

Hepatocellular carcinoma (HCC) is the most common primary liver cancer worldwide. The histological diagnosis of HCC is based on the classification of liver cancer cells inside and outside the tumour area. In particular, the analysis of the morphological and architectural features of cell nuclei allows the differentiation of neoplastic hepatocytes from healthy ones. Although traditional histopathological approaches remain crucial for the identification of cellular changes, they are time-consuming and exposed to high intra- and inter-observer variability. In addition, the intratumour heterogeneity of HCC contributes to making diagnostic analysis very challenging. In recent years, the classification of HCC has been improved by the advent of digital pathology (DP), which allows histological slides to be digitised, and the development of computer-aided diagnosis (CAD) systems, which exploit machine-learning algorithms to perform automatic image analysis and assist pathologists in visually inspecting liver tissue. This Ph.D. thesis proposes a CAD system integrated into the DP workflow for the automatic detection of HCC liver nucleus in confocal fluorescent images. Specifically, liver tissue samples were stained by immunofluorescence, and images were acquired using a confocal microscope implementing the z-stack technique. Subsequently, a CAD system based on a convolutional neural network with handcrafted feature injection was developed for HCC detection. The developed system achieved promising results in the classification of liver nucleus, and showed good adaptability to different diagnostic scenarios, including the histological diagnosis of lung and colon cancer, the classification of blood cells in haematological samples, and the distinction between nevi and melanomas in dermoscopic images.