This project involves creating and training convolutional neural networks on mammograms from the CBIS-DDSM dataset. The project started as a project for a course in machine learning I was taking at EPFL Extension School. That project involved doing binary classification of tiles extracted from the original images into normal and abnormal.

After finishing the course I was reviewing my work when I realized that any tile extracted from a positive image was predicted positive, regardless of whether the tile contained abnormalities of any sort. Further investigation revealed that the positive and negative images had some substantial differences in them, including pixel mean, pixel mins and maxes, standard deviation, and others. This meant that the model was using features unrelated to the actual abnormalities to make the predictions, rendering the entire model useless.

These problems were introduced by the inclusion of normal images from the DDSM dataset in addition to the abnormal CBIS-DDSM images. In order to force the network to focus on the actual abnormalities I decided to change the objective from binary classification to dense classification on a pixel level, effectively segmenting the image into normal and abnormal pixels.

This project is still ongoing and will be updated with any relevant developments. The source code is on GitHub and if anyone is interested or thinks they have anything to add, please feel free to fork the repository and/or send pull requests.

More information on this project is available here.

Last Updated: 2018-07-16 13:31

The source code is available here.