First blog

content/report/osre25/minyuan-20250812/first-blog.txt

@@ -0,0 +1,3 @@
+1. Research on multiple pathology images datasets and successfully detects a good dataset located at https://github.com/binli123/dsmil-wsi.
+2. Try different methods for reconstruction of pathology images including VAE and GAN initially, but I found that the reconstruction is not very clear. So I research and use diffusion in the latent layer of VAE and use super-resolution in the generation process to improve the quality of reconstructed images. In the end, I successfully built a comprehensive VAE system with the help of llm techniques to provide reasonably good reconstructed images based on the baseline in this repo: https://github.com/cvlab-stonybrook/Large-Image-Diffusion/tree/main.
+3. Parameter tuning process for the model architecture: • VAE Component Optimization: Systematically tuned the encoder-decoder architecture by experimenting with latent dimension sizes (512, 1024, 2048), adjusting the beta coefficient in KL divergence loss (0.1 to 1.0), and optimizing the number of residual blocks (4-8 layers) to achieve optimal balance between reconstruction fidelity and latent space regularization for pathology images. • Latent Diffusion Parameters: Fine-tuned the diffusion process by testing different noise scheduling strategies (linear vs. cosine), optimizing the number of denoising timesteps (50-1000), adjusting the U-Net learning rates (1e-4 to 1e-6), and calibrating the classifier-free guidance scale (1.0-7.5) to maintain stable training while preserving critical pathological features during generation. • Super-Resolution Integration: Optimized the upsampling pipeline by testing different scale factors (2x, 4x, 8x), balancing multiple loss functions (L1, perceptual, and adversarial losses with weight ratios 1.0:0.1:0.01), implementing progressive training schedules, and fine-tuning the feature extraction networks to ensure diagnostically relevant pathological details are preserved and enhanced at higher resolutions.

content/report/osre25/ucsc/DPE-Net – Disentangled Pathology Editing Network/20250626-MingyuanShao/index.md

@@ -0,0 +1,22 @@
+---
+title: "DPE-Net – Disentangled Pathology Editing Network"
+subtitle: ""
+summary:
+authors: 
+  - Mingyuan Shao
+tags: ["osre25"]
+categories: [AI, VAE, GAN, DIFFUSION]
+date: 2025-06-01
+lastmod: 2025-06-26
+featured: false
+draft: false
+
+# Featured image
+# To use, add an image named `featured.jpg/png` to your page's folder.
+# Focal points: Smart, Center, TopLeft, Top, TopRight, Left, Right, BottomLeft, Bottom, BottomRight.
+image:
+  caption: ""
+  focal_point: ""
+  preview_only: false
+---
+