Inception-Res-Net-v2, ResNet50, DenseNet121 and InceptionV3 convolution neural networks have been created with a small subset (1.4 million images) of the RadImageNet databases. The non-optimized algorithms have been shown to outperform ImageNet algorithms on medical imaging applications. Algorithms are available on request. Instructions for use of the algorithms are available on GitHub.
Improved Results - ROC
A model derived using transfer learning from a RadImageNet Inception V3 model significantly outperformed an ImageNet Inception V3 derived model in diagnosing COVID-19 on Chest CT (data drawn from our previous Nature Medicine publication). With little to no optimization of the models the large difference in performance using RadImageNet is clear in the ROC curve below.
Improved Interpretation - CAMs
Not only do models derived from RadImageNet provide improved results, the extracted features are more clinically relatable as can be seen in the gradient class activation map image below (brighter is more important). The RadImageNet model clearly highlights the areas of consolidation while the ImageNet model highlights unrelated soft tissue in the periphery of the image.