Since 2011, deep learning algorithms have been gaining more attention within the machine learning community, as their success rates on certain recognition tasks have been competitive with humans. In 2012, researchers from the University of Toronto trained a deep convolutional neural network (CNN) using the largest tagged dataset available, ImageNet, which consists of 15 million high-resolution images tagged in over 22,000 categories. They achieved object recognition top-1 and top-5 error rates of 37.5% and 15.3% respectively.Evaluation
We trained a convolutional neural network with 10,000 chest radiographs which were tagged as either normal (absence of any clinically relevant pathology) or abnormal based on the final radiology report. This was performed using a single 2 GB nVidia GTX 770 graphics processing unit (GPU) and an open-source deep learning software package (convnet). The performance of the neural network was tested using an untrained dataset consisting of 500 radiographs and characterized using receiver-operating curve analysis at different output probability thresholds. It achieved a maximum sensitivity of 95% with a corresponding specificity of 85%.Discussion
Advancements in machine learning have been possible due to improvements in computation power through the use of GPUs and the access to large quantities of data. The current success and future developments of these algorithms will have a profound effect on the interpretation of medical images. We have proven that with adequate data, these algorithms can be used to help automate and speed up medical diagnosis. With more data, we expect further improvement in performance. Furthermore, more experimentation is required to determine if diagnostic subcategories could be classified with this same approach.Conclusion
Convolutional neural networks can be trained using a modestly sized medical dataset to screen chest radiographs as normal or abnormal. Deep learning will play an integral role in advancing computer-aided diagnosis which will enhance and speed up the workflow of the radiologist. Further experimentation is required using larger datasets as well as different types of imaging studies.