In this repository we remake the results of this paper on breast image classification.
T. Araújo et al., “Classification of breast cancer histology images using Convolutional Neural Networks,” PLOS ONE, vol. 12, no. 6, p. e0177544, Jun. 2017.
The dataset used in this paper has been extended and is available as part of the ICIAR 2018 challenge. The challenge contains two tasks. We are concerned with the first task - classification of breast tissue images (2048x1536 pixels @ 20X). Once the dataset has been downloaded these images can be found in the subfolder called 'Photos'.
There are four classes (100 images per class):
- Benign
- InSitu Carcinoma
- Invasive Carcinoma
- Normal
image (below) from the challenge website
The general idea is to build a patch based classifier (CNN) and then aggregate the results (majority vote) to give an overall image class. The patch size used is 512x512.
Implementations of the CNNs are found in the folders 'keras_folder' and 'pytorch_folder'. I have achieved results comparable to the original paper with these implementations.