How wqli498/patholic works
PathoLIC compresses whole-slide pathology images with content-aware variable-rate learned compression, then evaluates reconstructed data and extracted features in downstream pathology tasks. The documented workflow runs WSI segmentation/patching/stitching, feature or attention extraction, compression training or checkpoint evaluation, bitstream decoding, and downstream validation. The tree also includes substantial local CompressAI and nnU-Net subsystems. Evidence is strongest for the README workflow and indexed imports; unsampled implementations and their internal wiring are represented conservatively.
Main entry points: main.py Workflow runner · vbr.py Variable-rate codec · decompression_and_extract_features.py Decode and extract · compression_model_validation.py Compression evaluation
Components 26
WSI preparation
- Workflow runner
main.py - Slide processing
create_patches_fp.py - WSI operations
wsi_core/WholeSlideImage.py
Learned compression
- Compression training
compression_model_training.py - Compression evaluation
compression_model_validation.py - Variable-rate codec
compressai/models/vbr.py - Entropy models
compressai/entropy_models/entropy_models.py - Latent codecs
compressai/latent_codecs/gaussian_conditional.py - Codec base models
compressai/models/base.py - Compressed bitstreams
Features and attention
- Patch labeling
patch_labeling.py - Decode and extract
decompression_and_extract_features.py - Feature encoders
gigapath/pipeline.py - TITAN encoder
mytitan/vision_transformer.py - CHIEF attention
CHIEF.py - Feature files
Downstream validation
- Task fine-tuning
finetune/main.py - Survival analysis
titan_survival_brca.py - ROI retrieval
uni_roi_retrieval_crc_vis.py - Patch classification
run_patch_classification.py - Nuclei segmentation
nnUNet/nnunetv2/inference/predict_from_raw_data.py - Segmentation plans
nnUNet/nnunetv2/experiment_planning/experiment_planners/default_experiment_planner.py
Model libraries
- Codec model catalog
compressai/zoo/image.py - Rate-distortion loss
compressai/losses/rate_distortion.py - Component registry
compressai/registry/torch.py
Other components
- Researcher
Connections 28
- Researcher → Workflow runner: runs workflow
- Workflow runner → Slide processing: prepares slides
main.py - Slide processing → WSI operations: uses slide operations
- Workflow runner → Patch labeling: extracts features
main.py - Workflow runner → Compression training: trains codec
main.py - Workflow runner → Compression evaluation: evaluates codec
main.py - Compression training → Variable-rate codec: optimizes
- Compression evaluation → Variable-rate codec: evaluates
- Variable-rate codec → Entropy models: uses
compressai/models/vbr.py - Variable-rate codec → Codec base models: extends
compressai/models/vbr.py - Entropy models → Latent codecs: provides coding models
- Workflow runner → Decode and extract: invokes extraction
main.py - Decode and extract → Compressed bitstreams: reads and decodes
- Compression evaluation → Compressed bitstreams: writes
- Decode and extract → Feature files: writes features
- Patch labeling → Feature encoders: extracts embeddings
- Decode and extract → TITAN encoder: supports TITAN features
- Workflow runner → CHIEF attention: supports CHIEF attention
- Workflow runner → TITAN encoder: supports TITAN attention
- Researcher → Task fine-tuning: runs task training
- Researcher → Survival analysis: validates survival
- Researcher → ROI retrieval: runs retrieval
- Researcher → Patch classification: runs classification
- Researcher → Nuclei segmentation: runs segmentation
- Segmentation plans → Nuclei segmentation: plans inference
- Codec model catalog → Codec base models: uses model base
compressai/zoo/image.py - Variable-rate codec → Rate-distortion loss: trained with
- Codec model catalog → Component registry: registers models