Update MAISI README with Quality Evaluation Results

generation/maisi/README.md

@@ -8,6 +8,29 @@ This example demonstrates the applications of training and validating NVIDIA MAI
 
 ## Example Results and Evaluation
 
+We retrained several state-of-the-art diffusion model-based methods using our dataset. The results in the table and figure below show that our method outperforms previous methods on an unseen dataset ([autoPET 2023](https://www.nature.com/articles/s41597-022-01718-3)). Our method shows superior performance to previous methods based on all [Fréchet Inception Distance (FID)](https://papers.nips.cc/paper/2017/hash/8a1d694707eb0fefe65871369074926d-Abstract.html) scores on different 2D planes. Here we compared the generated images with real images of size 512 &times; 512 &times; 512 and spacing 1.0 &times; 1.0 &times; 1.0 mm<sup>3</sup>.
+
+<div align="center">
+
+| Method | FID (XY Plane) ↓ | FID (YZ Plane) ↓ | FID (ZX Plane) ↓ | FID (Average) ↓ |
+|--------|:----------------:|:----------------:|:----------------:|:---------------:|
+| [DDPM](https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf)   |      18.524       |      23.696      |      25.604      |      22.608     |
+| [LDM](https://openaccess.thecvf.com/content/CVPR2022/papers/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.pdf)    |      16.853       |      10.191      |      10.093      |      12.379     |
+| [HA-GAN](https://ieeexplore.ieee.org/document/9770375) |      17.432       |      10.266      |      13.572      |      13.757     |
+| MAISI  |       3.301       |       5.838      |      9.109       |      6.083      |
+
+**Table 1.** Comparison of Fréchet Inception Distance scores between our foundation model and retrained baseline methods<br>using the unseen public dataset [autoPET 2023](https://www.nature.com/articles/s41597-022-01718-3) as the reference.
+
+</div>
+
+<div align="center">
+
+![Generated image examples](https://developer-blogs.nvidia.com/wp-content/uploads/2024/06/generated-medical-image-method-comparison-1.png)
+
+**Figure 1.** Qualitative comparison of generated images between baseline methods<br>(retrained using our large-scale dataset) and our method.
+
+</div>
+
 ## MAISI Model Workflow
 The training and inference workflows of MAISI are depicted in the figure below. It begins by training an autoencoder in pixel space to encode images into latent features. Following that, it trains a diffusion model in the latent space to denoise the noisy latent features. During inference, it first generates latent features from random noise by applying multiple denoising steps using the trained diffusion model. Finally, it decodes the denoised latent features into images using the trained autoencoder.
 <p align="center">