Using artificial intelligence to automatically detect breast cancer

Methods

Summary

The tissue is prepared as usual by the lab technician (slicing and staining). After the pathologist has established the diagnosis the sample is put in a whole slide scanner that generates an image of the specimen.

The specimen is labelled manually and put into the image database. The image database is divided in a training set, a validation set and a test set. The test set is used only at the end. Every step (called epoch) first the training set is passed throug the network and the results are recorded. From the difference of the (known) true and calculated results a correction of the parameters (called weights) of the CNN is calculated. After all images of the training set have been used the validation set is passed thru the network but no adjustments of the weights are made. The accuracy of the results are recorded. If not sufficient, the training is repeated with the training set as described above.

After having reached the desired accuracy the test set is used as input for the network and the final result is displayed.

To predict the presence of tumour in new images the image under consideration is sliced into small snippets and these are used as input data of the neural network. If classified positive the corresponding area of the original image is coloured.

The programs necessary have been written in Python using the Keras/Tensorflow libraries.

For more theoretical background see:

Backpropagation algorithm

Challenges

Currently the greatest obstacle is the lack of a powerfull graphics card. Training currently is too slow to be efficient. In order to find the best CNN many different parameter combinations have to be tested.

Pre Analysis Plan

We measure accuracy and loss of the results. These are the usual metrics used for neural networks.