Predicting Cancer metastates

The Problem

Dataset: Biopsy images at 8 level(0-7) magnification, with corresponding masks

Goal: Assisting doctors by detecting regions that likely have metastatis (at level 7)

Methodology(compelte work done on google colab):

1. Extract data:

   a. Extract few images and corresponding masks by

   b. Extract cropped patches from slides from levels 5, 6, 7

   c. Generate labels by looking at mask of the center region

2. Train model

3. Test on one image

Built-in Functions

• read_slide(slide, x, y, level, width, height, as_float=False) - Reads slides
• positive_samples_from_image(image_slide, mask_slide, number=100, level=5, patch_size=(100, 100)) - outputs 'number' of samples with tumor for a given patch size
• negative_samples_from_image(image_slide, mask_slide, number=100, level=5, patch_size=(100, 100)) - outputs 'number' of samples without tumor for a given patch size
• test_image(image_slide, model= model_vgg, level=7, patch_size=(100, 100))- output predicted mask for given image

Dataset used

• xpos, xng, ypos, yneg - .npy files containing positive and negative patches extracted from 5 slides at (5, 6, 7) levels. y are the associated labels

Models

• vgg_model.h5 (best performing, with 0.96 ROC AUC)
• inception_model.h5

Model Results (test accuracy, precision, recall, roc auc)

• **VGG: 90%, 87%, 87%, 0.956
• **Inception: 84%, 80%, 75%, 0.905

Useful links

• Sample data: https://drive.google.com/drive/folders/1rwWL8zU9v0M27BtQKI52bF6bVLW82RL5
• Youtube Video: https://youtu.be/eftMWQRCPRg
• Research Paper: https://arxiv.org/pdf/1703.02442.pdf