Automated detection of liver steatosis in ultrasound images using convolutional neural networks

Ultrasound imaging is the most commonly applied imaging modality for the diagnosis of fatty liver disease. It is considered malignant if there is more than 5% of fatty hepatorenal steatosis. The classical methods to classify liver steatosis usually involve experienced physicians or radiologists to identity them. In this work, we introduce a Convolutional Neural Network (CNN) based approach to classify the malignant and benign fatty livers from ultrasound images. The pre-trained network of Inception Resnet which is initially trained on the ImageNet dataset is used for transfer learning on B-mode ultrasound liver images for classification. We used the open-source ultrasound liver dataset of 55 patients with 10 image sequences for each making a total of 550 images with 170 benign and 340 malignant samples. Since the dataset size is small for training, we have applied various data-augmentation techniques and have employed the transfer-learning approach using Inception ResNet architecture. We were able to achieve, using our approach, a very high classification accuracy of 98.48%, whereas the area under the curve of the classical hepatorenal index method is 0.959 and the Gray-level co-occurrence algorithm is 0.893.