Deep Learning Based Steatosis Quantification of Liver Histopathology Images Using Unsupervised Feature Extraction


KARAGÖZ M. A., KARABOĞA D., AKAY B., BAŞTÜRK A., Nalbantoglu O. U.

2nd International Conference on Computing and Machine Intelligence, ICMI 2022, İstanbul, Turkey, 15 - 16 July 2022 identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/icmi55296.2022.9873795
  • City: İstanbul
  • Country: Turkey
  • Keywords: deep learning, steatosis quantification, transfer learning, UNet
  • Erciyes University Affiliated: Yes

Abstract

© 2022 IEEE.Steatosis quantification is an essential issue for accurate diagnosis and donor transplantation. However, manually quantification processes of steatosis by a pathologist have some difficulties because of time-consuming and tiring processes that can vary in inter and intra-experts. In recent years, deep learning studies have emerged with promising performance on steatosis quantification. On the other hand, deep learning models require a large amount of data, yet the steatosis dataset is insufficient for deep models. Thus, we propose deep learning model consisting of two steps that showed high performance even on a small number of steatosis datasets. The first step is unsupervised feature extraction with UNet. The second step is classification by using extracted features as an input for classification models. ResNet-50, EfficientNet B1 and MobileNetV2 networks are used for classification. As a result, the proposed deep models enable fully automated steatosis quantification with high AUC.