PyTorch implementation for Pathology Image Restoration via Mixture of Prompts .

This repo contains:

• MoP
  • Overview
  • Training of defocus encoder
    • Dataset Format
    • Usage
  • Trainnig of prompt restorer
  • Training of P-former
  • Training of P-diffusion
  • Supplementary annotations and test code on 3DHistech

• Each folder contains the original and fused images from multiple fields of view for a single WSI.

  • xxx represents the id of the field.
  • (xxx_01.jpg, xxx_02.jpg): Represents the low-quality images from different focal planes.
  • (Z_xxx.jpg): Represents the corresponding fused image of a field.

• The training, validation, and test sets are recorded in separate .txt files.

• Each line in the .txt file contains information of a cropped image:

  • The image path
  • The CTF value
  • Distance from the optimal plane

Download the weights of Resnet-18 before training.

Train the defocus encoder by:

cd defocus_encoder
python train.py

Download the weights of various pathology foundation models before training.

Train the prompt restorer by:

cd prompt_restorer
python train.py

Only for inference:

python infer.py

Supplementary annotations are provided 3D_anno.csv. 0 represents normal cells and 1 represents abnormal cells.