An unsupervised transfer learning model based on convolutional auto encoder for non-alcoholic steatohepatitis activity scoring and fibrosis staging of liver histopathological images

Non-alcoholic fatty liver disease (NAFLD) is one of the most frequent chronic liver diseases worldwide. Non-alcoholic steatohepatitis (NASH) is a progressive type of NAFLD that may cause cirrhosis, hepatocellular carcinoma, or almost mortality. Therefore, early diagnosis of the NASH is crucial. NASH is scored using the main histopathological features: ballooning, inflammation, steatosis, and fibrosis. The diagnosis of NASH by pathologists is time-consuming and can vary subjectively. On the other hand, several studies have reported deep learning approaches to enable fully automated NASH scoring. However, these studies suffer from limited labeled and imbalanced datasets. The purpose of this study to achieve fully automated NASH scoring with deep learning models that overcome data limitations. This study proposes an unsupervised transfer learning model for NASH scoring and fibrosis staging on a small-size dataset within two steps. In the first step, Convolutional Auto Encoder (CAE) is utilized as a deep feature extractor in an unsupervised manner during reconstruction. The second step is fine-tuning for classification consisting of the CAE encoder set as initial layers combined with fully connected layers and softmax. The proposed unsupervised transfer learning model is evaluated on a public NASH dataset. We compare the performance of the proposed network (CAE+classifier) with transfer learning models including Inception-v3, VGG16, ResNet-50. The proposed model has 94.87% AUC for ballooning, 89.47% AUC for inflammation, 96.15% AUC for steatosis, and 93.18% AUC for fibrosis. The proposed model is superior to transfer learning models (Inception-v3, VGG16, ResNet-50) with less parameter size and low computational complexity on a small NASH dataset.