And end-to-end artificial intelligence model accurately diagnosing hepatocellular carcinoma on computed tomography

Abstract

Introduction: Hepatocellular carcinoma (HCC) is the fourth leading cause of cancer death worldwide. Diagnosis often requires highly characteristic dynamic patterns on contrast-enhancing cross-sectional imaging. The Liver Imaging Reporting and Data System (LI-RADS), used to interpret radiological findings related to HCC, is often limited by a lack of definitive diagnosis, resulting in treatment delays. We developed an artificial intelligence deep learning algorithm applying state-of-the-art convoluted neural network (CNN) architecture in detecting and classifying liver lesions on computed tomography (CT). Aims & Methods: We retrieved archived tri-phasic contrast liver CT images and clinical information from different medical centers in Hong Kong and Shenzhen. We followed American Association for the Study of Liver Diseases recommendations for HCC diagnosis and employed LI-RADS classification in lesion categorization, with diagnosis validated by a clinical composite reference standard based on patients’ outcomes over the subsequent 12 months. Each liver lesion was manually contoured and labelled with diagnostic ground-truth. After data augmentation and lesion segmentation over all CT phases, we applied different architectural networks for deep machine learning (NVIDIA Tesla V100 GPUs, Dell Technologies, Singapore), including a densely-CNN consisting of five dense blocks, each with 64 convolutional layers, followed by a fully-connected layer leading to a 186-layer network, with softmax as the activation function for classification. Results: As of April 2020, we have retrieved and contoured 1,334 scans with 2,631 liver lesions. The cohort’s mean age was 59.3±13.6 years, 63.3% male. Mean lesion size was 36.6±44.5 mm, with 871 lesions (33.1%) confirmed as HCC. Heat maps were generated and superimposed on original images based on the model’s prediction of likelihood of lesion location. Such heat map hot-zones appropriately covered the location of lesions in CT images, allowing the detection of existing lesions without further annotated contours. An interim analysis of 1,693 lesions, divided in a 7:3 ratio into training and testing sets, found the densely-CNN to achieve a diagnostic accuracy of 97.0%, negative predictive value (NPV) 99.0%, positive predictive value (PPV) 92.4%, sensitivity 97.3% and specificity 96.9% in the binary classification of HCC. This was compared to the diagnostic accuracy of 86.2%, NPV 86.0%, PPV 86.5%, sensitivity 84.6% and specificity 87.8% via LI-RADS. Over 115 million radiological parameters were optimized in the development of our deep learning model. Conclusion: Our deep learning model achieved a high diagnostic accuracy based on an end-to-end densely-CNN in detecting and classifying HCC vs. non-HCC. Further confirmatory analyses will be performed in an external validation group with >1,000 scan images. Artificial intelligence has potential in improving the diagnostic capabilities of HCC and prevent delays in treatment. References: Supported by the Innovation and Technology Fund, the Government of the HKSAR (ITS/122/18FP); and United Ally Research Limited, a subsidiary of Hong Kong Sanatorium and Hospital Limited.