Deep Learning-Based Universal Expert-Level Recognizing Pathological Images of Hepatocellular Carcinoma and beyond

Purpose: Human based medical-image interpretation always falls into the predicament between specialized practitioners and expanding medical imaging. We aim at developing a diagnostic tool for pathological-image classification by using transfer learning that can be applied to diverse tumor types. Experimental Design: In this study, images were retrospectively collected and prospectively analyzed using machine learning. Microscopic images of liver tissue that show or do not show hepatocellular carcinoma were used to train and validate a classification framework based on convolutional neural network. To evaluate the universal classification performance of the artificial-intelligence (AI) framework, histological images from colorectal tissue and breast were also collected. Training and validation set of images were collected from Xiamen Hospital of Traditional Chinese Medicine whereas test set of images were collected from Zhongshan Hospital Xiamen University. Results: Accuracy, sensitivity, and specificity were reported and compared to human image interpretation and other AI image classification systems such as AlexNet and GoogLeNet. For the test dataset, sensitivity, specificity, and area under the curve of the AI framework were 99.1%, 98.0%, and 0.960, respectively. In human-machine comparisons, the accuracy of the AI framework was 98.5%, while the accuracy of human experts fluctuated between 93.0% and 95.0%. Based on transfer learning, the AI framework accuracy for colorectal carcinoma, breast invasive ductal carcinoma, were 96.8%, and 96.0%, respectively. Conclusions: The performance of the proposed AI framework in classifying histological images with hepatocellular carcinoma is comparable to the classification by human experts. With limited training, the proposed AI framework has potential universality in histological image classification.