Accuracy in classification of liver lesion ultrasound using artificial intelligence

Abstract

B-mode ultrasound imaging is the standard method for hepatocellular carcinoma (HCC) screening. If a lesion has a unique appearance, a conventional detector can be applied. However, if the lesion shares its appearance with other lesion, a large training dataset is required, which may be unavailable. Therefore, a two-stage method is proposed. In the first stage, lesions are detected but not differentiated into any particular class. The detected lesions are then classified in the second stage using a conventional convolutional neural network (CNN).

The aims of the dissertation are to design an artificial intelligent system model and to investigate the most accurate deep learning structure for hepatic lesion classification using two-stage model.

Even though the cysts and hepatic vessels are both anechoic pattern and present black oval in ultrasound images, cysts have unique artifacts that present posterior acoustic enhancement. By including the artifacts in the detection, conventional detectors can be applied. In the study, Region-based convolutional neural networks, R-CNN, with Residual Network-50, ResNet-50, as the backbone was applied for the detection of 615 hepatic cysts. The system was evaluated by five-fold cross validation. The result indicated that the addition of artifacts led to better detection in term of accuracy, reduction of false positives and false negatives.

Hemangioma (HEM) and HCC share lots of sonographic appearance. There is no unique artifact to differentiate these two lesions. A two-stage method is applied. In the first stage, the detector is trained to identify HEM and HCC like lesions. In the second stage, the classifier is applied to differentiate lesions into typical HEM, atypical HEM and HCC. Since atypical HEM and HCC display the same sonographic appearance, both lesions cannot be differentiated solely by B-mode ultrasound images. They require further CT or MR investigation, so it is unnecessary to accurately differentiate these two lesions during screening. The study showed that grouping HCC and atypical HEM into the same one led to the increase of HCC recall of the detector from 0.64 to 0.68. The application of the two-stage method in place of detector only method improved the recall from 0.68 to 0.72. The recall rate was comparable to the detector only method that was trained by a much larger database and used more relax criterion (0.5 and 0.2 intersection over union (IoU) for the correct detection in the proposed two-stage and the detector only methods, respectively).