How hoomesh0407/oral-cancer-prediction works
This repository is an oral-pathology image classification prototype: a web API accepts a slide image, predicts a differentiation grade, and returns a report; separate scripts train and compare four-class and three-class models and generate explanations. The main API path is visible in excerpts, while some model-loading and artifact wiring is not. Training and evaluation scripts are therefore shown as independently evidenced branches rather than an assumed end-to-end pipeline.
Main entry points: app_lora.py Prediction API · explain_lora.py LoRA explanations · extract_embeddings.py Embedding extraction · extract_embeddings_3class.py 3-class embeddings
Components 23
Prediction product
- Web interface
index.html - Prediction API
app_lora.py - Image preparation
app_lora.py - Grade prediction
app_lora.py - Clinical report
app_lora.py - Premium interface
premium.html - Doctor interface
doctor.html
Model development
- Embedding extraction
extract_embeddings.py - 3-class embeddings
extract_embeddings_3class.py - Classifier training
train_classifier.py - LoRA training
train_lora.py - 3-class LoRA
train_lora_3class.py - DenseNet training
train_densenet_3class.py - Frozen classifier
train_frozen_3class.py - Fusion training
train_fusion_3class.py
Model evaluation
- Baseline comparison
baseline_comparison.py - Overfit checks
overfit_check_all.py - Overfit check
check_overfitting.py
Interpretability
- LoRA explanations
explain_lora.py
Other components
- Clinician
- MedSigLIP model
- Slide images
- Model artifacts
Connections 31
- Clinician → Web interface: opens
- Web interface → Prediction API: submits images
- Prediction API → Image preparation: cleans image
app_lora.py - Image preparation → Grade prediction: prepares pixels
app_lora.py - Grade prediction → Clinical report: produces grade
app_lora.py - Clinical report → Prediction API: returns result
app_lora.py - Prediction API → Premium interface: serves optionally
- Prediction API → Doctor interface: serves doctor route
- Clinician → Prediction API: uploads slide
app_lora.py - Grade prediction → MedSigLIP model: runs vision model
- Slide images → Embedding extraction: reads images
extract_embeddings.py - MedSigLIP model → Embedding extraction: extracts features
extract_embeddings.py - Embedding extraction → Model artifacts: saves embeddings
- Slide images → 3-class embeddings: reads images
extract_embeddings_3class.py - MedSigLIP model → 3-class embeddings: extracts features
extract_embeddings_3class.py - 3-class embeddings → Model artifacts: saves tensors
extract_embeddings_3class.py - Model artifacts → Classifier training: loads embeddings
train_classifier.py - Model artifacts → Frozen classifier: loads embeddings
train_frozen_3class.py - Slide images → LoRA training: loads training images
train_lora.py - MedSigLIP model → LoRA training: fine-tunes vision tower
train_lora.py - Slide images → 3-class LoRA: loads training images
train_lora_3class.py - MedSigLIP model → 3-class LoRA: fine-tunes vision tower
train_lora_3class.py - Slide images → DenseNet training: loads training images
train_densenet_3class.py - Model artifacts → Fusion training: loads feature tensors
train_fusion_3class.py - Slide images → Baseline comparison: evaluates images
baseline_comparison.py - Slide images → Overfit checks: checks model splits
overfit_check_all.py - Slide images → Overfit check: checks overfitting
- Slide images → LoRA explanations: loads example slides
explain_lora.py - MedSigLIP model → LoRA explanations: computes activations
explain_lora.py - Model artifacts → LoRA explanations: loads adapter and head
explain_lora.py - LoRA explanations → Model artifacts: saves overlays
explain_lora.py