Non-invasive multi-channel deep learning convolutional neural networks for localization and classification of common hepatic lesions

Introduction
Machine learning techniques, especially convolutional neural networks (CNN), have revolutionized the spectrum of computer vision tasks with a primary focus on supervised and labelled image datasets. We aimed to assess a novel method to segment the liver from the abdomen computed tomography (CT) image using the CNN network, and to train a unique method to locate and classify liver lesion pre-histological findings using multi-channel deep learning CNN (MDL-CNN).

Material and methods
The post-contrast CT images of the liver with a resolution of 0.625 mm were chosen for the study. In a random method, 50 examples of each hepatocellular carcinomas, metastases tumours, haemangiomas, hepatic cysts were chosen and evaluated.

Results
The dice score quantitatively analyses the similarity of segmentation results with the training dataset. In the first CNN model for segmenting the liver, the dice score was 96.18%. The MDL-CNN model yielded 98.78% accuracy in classification, and the dice score for locating liver lesions was 95.70%. Additionally, the performance of this model was compared to various other existing models.

Conclusions
According to our study, the machine learning approach can be successfully implemented to segment the liver and classify lesions, which will help radiologists impart better diagnosis.