/Detecting-Cancer-on-Gigapixel-Images

Columbia University Applied Deep Learning Project: Detecting Cancer Metastases on Gigapixel Pathology Images

Primary LanguageJupyter Notebook

Detecting Cancer Metastases on Gigapixel Pathology Images

Columbia University Applied Deep Learning Course Project (Fall 2019)
Author: Yingxiang Chen
Columbia Uni: yc3526
Video Demo: https://youtu.be/h6wJMuvgd4M

Objective

Current Situation:

  • Microscopic examination of lymph nodes is crucial in breast cancer staging
  • Currently, the manual process requires highly skilled pathologists
  • The process is fairly time-consuming and error-prone, particularly for lymph nodes with either no or small tumors

So I want to follow the strategy in the paper and utilize the deep learning models & techniques to relieve the workload of physicians by creating a workflow to detect and locate tumor pixels in the images and offer automatic second opinions.

Dataset

  • Raw Data: 21 Gigapixel Pathology Images, each has a tumor slide and a corresponding mask from The CAMELYON16 challenge.
  • Trainset: 8000 (4000 each at two different zoom levels) image patches sampled from 16 Gigapixel Pathology Images.
  • Validation set: 1600 (800 each at two different zoom levels) image patches sampled from 2 Gigapixel Pathology Images.
  • Test set: 3 Gigapixel Pathology Images.

Image augmentation

Adopt similar augmentation strategies discussed in the paper.

  • Use Keras ImageDataGenerator to augment data

    • Horizontal_flip
    • Vertical_flip
    • Rescale
    • Width_shift
    • Height_shift
    • Rotation
  • Use TensorFlow image random function to augment data

    • Random brightness
    • Random saturation
    • Random hue
    • Random contrast

Model

  • Transfer Learning: Used two pre-trained inception v3 models on Imagenet to speed up the training process.
  • Global Pooling: Applied GlobalAveragePooling layer after the inception model to significantly reduce parameters.

img

Result

  • The result on slide 075

img

  • THe result on slide 091

img

  • The result on slide 096

img

Final Deliverable