AdvDINO Pipeline

Workflow order:

1. Preprocessing → Prepare and process mIF tiles
2. AdvDINO → Train domain-adversarial self-supervised embeddings
3. Analysis
   • ABMIL_OS → Perform survival prediction with attention-based MIL
   • Clustering → Perform unsupervised clustering and analyze cluster phenotypes

Preprocessing

This folder contains scripts for preparing imaages for downstream analysis.
The preprocessing workflow includes three main steps:

1. Downsampling WSIs

   • Script: downsample_images.py

2. Generating Tissue Masks

   • Script: save_masks.py

3. Processing Tiles

   • Script: process_tiles.py

---

Usage

1. Downsample WSIs

python downsample_images.py --ipt_dir /path/to/images --opt_dir /path/to/downsampled --start 0 --stop 500

2. Generate Masks

python save_masks.py --ipt_dir /path/to/downsampled --opt_dir /path/to/masks

3. Process Tiles

python process_tiles.py --ipt_dir /path/to/images --mask_dir /path/to/masks --opt_dir /path/to/tiles --start 0 --stop 500

---

AdvDINO

This folder contains scripts for domain-adversarial self-supervised training and tile-level embedding extraction for downstream analysis. The AdvDINO workflow includes two main steps:

1. Training

   • Script: train.py

2. Inference (Generate Embeddings)

   • Script: inference.py

---

Usage

1. Training

python -m torch.distributed.launch -nproc_per_node=num_gpus dinov2/train/train.py --config-file=/path/to/configs --output-dir=/path/to/outputs train.dataset_path=mIFDataset:root=/path/to/tiles

2. Inference (Generate Embeddings)

python dinov2/eval/tile_inference.py --config-file=/path/to/configs --pretrained-weights=/path/to/weights --dataset_str=mIFCoordDataset --input_dir=/path/to/images --output_dir=/path/to/embeddings --start=0 --stop=500 --batch_size=128 --num-workers 8 --device_num=1

---

ABMIL_OS

This folder contains scripts for attention-based multiple instance learning (ABMIL) to perform survival prediction. The workflow includes three main steps:

1. Generate Split Folds

   • Script: final_gen_splits_primary.py
   • Generates train/val/test splits and folds for cross-validation for survival analysis.

2. Compute Quartiles

   • Script: get_quartiles.py
   • Computes survival time quartiles necessary for updating the configuration file for ABMIL.

3. Run ABMIL-OS Inference

   • Script: main.py
   • Performs survival prediction using tile embeddings (.h5 files) from AdvDINO inference.

---

Usage

1. Generate Split Folds

python final_gen_splits_primary.py --labels_path /path/to/labels --embeddings_dir /path/to/embeddings --num_folds 5 --output_dir /path/to/folds --seed 42

2. Compute Quartiles for Config

python get_quartiles.py --labels_path /path/to/labels --ipt_dir /path/to/embeddings 

3. Run ABMIL-OS for Survival Prediction

python main.py --config configs/default_config.yaml --embed_dim 1024 --embeddings_dir /path/to/embeddings --learning_rate 1e-4 --epochs 40 --data_source mif --cross_validation

---

Clustering

This folder contains scripts for unsupervised clustering of tile embeddings to identify proteomic phenotypes. The clustering workflow includes three main steps:

1. Combine Tile Embeddings

   • Script: combine_tile_embeddings.py
   • Combines all tile-level .h5 embeddings into a single matrix for clustering.
   • subsample_combine_tile_embeddings.py can be used to combine a subset of embeddings for faster exploratory clustering.

2. Perform Subsampled Clustering

   • Script: mIF_clustering.py
   • Performs Leiden clustering on subsampled tile embeddings to identify phenotypic clusters.

3. Cluster Attribution (Optional)

   • Script: clustering_attribution.py
   • Assigns cluster labels to the remaining tiles based on the nearest cluster from the subsampled clustering.
   • Produces full-slide cluster assignments and can generate summary outputs for downstream analysis.

---

Usage

1. Combine Tile Embeddings

python combine_tile_embeddings.py --labels_path /path/to/labels --embeddings_path /path/to/embeddings --output_path /path/to/output

(Optional: Subsample for faster clustering)

python subsample_combine_tile_embeddings.py --labels_path /path/to/labels --embeddings_path /path/to/embeddings --output_path /path/to/output --multiplier 0.2

2. Perform Clustering

python mIF_clustering.py --input_path /path/to/combined/embeddings --sample_tile_count 2000000  --n_pcs 128 --resolution 2.0 --n_neighbors 250

3. Attribute Remaining Tiles to Clusters (Only for Subsampled)

python clustering_attribution.py --main_cluster_path /path/to/combined/embeddings --target_name remainder_sub0.2_embeddings --ref_name subsampled_clustering --combined_name combined_name --cluster_column leiden_2.0

---