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

PURPOSE: 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.