This project implements a deep learning pipeline for detecting metastatic cancer in small histopathology image patches. It focuses on handling .tif images efficiently, training a binary classifier, and generating predictions for Kaggle’s Histopathologic Cancer Detection competition.

• Task: Binary classification — predict if the center 32×32 px region of a 96×96 px patch contains tumor tissue.
• Input: .tif pathology image patches
• Output: Probability of tumor presence
• Metric: ROC AUC

• All images are provided as .tif files.

• Labels are in train_labels.csv with columns:

  • id: image ID, mapping to train/{id}.tif
  • label: 1 for tumor, 0 for normal

Example:

• Custom loaders handle 8/16-bit .tif images and normalize them to float32 [0,1].
• Automatically ensures RGB output.

• Loads and checks label distribution.
• Validates that all .tif files exist.
• Visualizes sample positive and negative patches.

• Uses Albumentations for strong data augmentation.
• Custom Dataset class reads .tif files and converts them to tensors.

• EfficientNet-B0 (from timm) is used as the backbone.
• Outputs a single logit for binary classification.

• Stratified K-Fold cross-validation.
• Binary cross-entropy loss and ROC AUC metric.
• Saves the best checkpoint per fold.

• Test-time augmentation (TTA) improves performance.
• Combines predictions from multiple folds.
• Generates a final submission.csv for Kaggle.

• Larger backbones or higher resolution inputs
• Stain-aware normalization or augmentation
• Class weighting or focal loss
• Semi-supervised learning with pseudo-labeling
• Model ensembling

• ✅ submission.csv — ready for Kaggle submission
• ✅ Full .ipynb training & inference notebook
• ✅ Example visualizations and metrics

• Kaggle: Histopathologic Cancer Detection
• timm
• Albumentations