prashantksharma / FourthBrainBreastCancer

This is our final project for Fourth Brain

FourthBrainBreastCancer

This is our final project for Fourth Brain

Cancer Map

Libraries

Stage	Libraries
Prototyping	Pandas, Numpy
WSI Tools	OpenSlide
DL	Tensorflow
API	FastAPI
Front End	Dash?

Sources

White Papers

1. A Comprehensive Review for Breast Histopathology Image Analysis Using Classical and Deep Neural Networks
2. A Fast and Refined Cancer Regions Segmentation Framework in Whole-slide Breast Pathological Images
3. Assessment of Breast Cancer Histology using Densely Connected Convolutional Networks
4. A Unified Framework for Tumor Proliferation Score Prediction in Breast Histopathology
5. Deep Learning for Identifying Metastatic Breast Cancer
6. Detecting Cancer Metastases on Gigapixel Pathology Images - Google 2017t
7. Multi-Stage Pathological Image Classification using Semantic Segmentation

Other Works

Current development / How to use :

base_directory/dataset_folder

base_directory
├── dataset_folder
    ├── training
    │   ├── lesion_annotations
    │   │   └── tumor_001.xml
    │   ├── normal
    │   │   └── normal_001.tif
    │   └── tumor
    │       └── tumor_001.tif
    │
    └── testing
        ├── lesion_annotations
        │   └── test_001.xml
        └── images
            └── test_001.tif

Implemented so far:

• Generate_tiles.py script :

Takes the normal (negative), tumoral (positive) WSIs and corresponding lesion annotations (xml). And stores the tiles into hdfs files stored in a destination folder.

Note: During the next stage, we will generate augmented tiles that will be used in our training model. In order to add randomness, the tiles generated with generate_tiles.py should be larger than the ones used in read_tiles.py In our case, we generate tiles of 312 x 312. Later that tile will be randomly cropped into a 256 x 256 tile.

• Read_tiles.py prepares data for the training and validation process. The generator can be directly plugged into a model.fit() call The data augmentation and color normalization is made at that level

Next steps:

• Upstream WSI cleaning in generate_tiles.py to improve the quality of the training set generated
• Test dataset generation