Fig. 7:

When solving problems in volumetric medical modalities, such as CT and MRI images, 3D volume-based approaches consistently offer superior performance than 2D slice-based approaches empowered by transfer learning. We conduct statistical analyses (circled in blue) between the highest performance achieved by 3D and 2D solutions. Training 3D models from scratch does not necessarily outperform their 2D counterparts (see NCC). However, training the same 3D models from Genesis Chest CT outperforms all their 2D counterparts, including fine-tuning Models ImageNet as well as fine-tuning our Genesis Chest X-ray and Genesis Chest CT 2D. It confirms the effectiveness of Genesis Chest CT in unlocking the power of 3D models. In addition, we have also provided statistical analyses between the highest and the second highest performances achieved by 2D models, finding that Models Genesis (2D) offer equivalent performances (n.s.) to Models ImageNet in four out of the five applications.