How sacmehta/ynet works
Project type and purpose
• Y-Net is a deep-learning research codebase for joint tissue segmentation and diagnostic classification of breast biopsy images. It implements a two-stage training pipeline (segmentation only → joint segmentation+classification) plus a segmentation evaluation module.
File structure
• Top-level directories:
• stage1/ – code for training the segmentation network (branch 1)
• stage2/ – code for training the joint segmentation+classification network (branch 2)
• seg_eval/ – code for producing and evaluating segmentation masks with the trained Y-Net
• images/ – example input/output illustrations
• Within each stage:
• Data ingestion & augmentation: loadData.py, Transforms.py, DataSet.py
• Model definition: Model.py
• Loss & metrics: Criteria.py, IOUEval.py
• Training orchestration: main.py
• Visualization: VisualizeGraph.py
• Pretrained weights folder (pretrained_models_st1, pretrained_model_st2)
README for additional insights
• Two-stage training: stage1 trains segmentation branch, stage2 initializes from stage1 weights and jointly trains for segmentation + classification.
• seg_eval runs the trained model on new whole-slide images to produce segmentation masks.
• Dependencies: Python 3.6.2, PyTorch 0.2.0_4, OpenCV 3.3.0
Main components
a. Input Data
• Raw WSI patches (RGB) + ground-truth masks (annotations)
b. Data Pipeline (common to both stages)
• loadData.py → DataSet → Transforms → DataLoader
c. Stage 1: Segmentation Training
• Model Definition (U-Net-style branch)
• Loss & Metric (e.g., cross-entropy, IOU evaluation)
• Visualization (graph, training curves)
• Output: pretrained segmentation weights
d. Stage 2: Joint Segmentation + Classification Training
• Model Definition (Y-Net combining segmentation and classification heads)
• Initialization from stage1 weights
• Losses: segmentation + classification
• Metrics & Visualization
• Output: final joint model weights
e. seg_eval Module
• Load final model weights
• Inference engine (Eval_YNet.py)
• Produce masks on new WSI sets
• Compute evaluation metrics (IOU, etc.)
f. External Libraries & Frameworks
• PyTorch
• OpenCV
• Python runtime
Architectural patterns & design principles
• Modular “pipeline” stages (stage1, stage2, seg_eval)
• Reusable data ingestion & augmentation modules
• Clear separation of concerns: data, model, training loop, evaluation, visualization
• Checkpointing: intermediate and final model artifacts