Breast mass segmentation from mammograms with deep transfer learning

Mammography is an x-ray imaging method used in breast cancer screening, which is a time consuming process. Many different computer assisted diagnosis have been created to hasten the image analysis. Deep learning is the use of multilayered neural networks for solving different tasks. Deep learning methods are becoming more advanced and popular for segmenting images. One deep transfer learning method is to use these neural networks with pretrained weights, which typically improves the neural networks performance. In this thesis deep transfer learning was used to segment cancerous masses from mammography images. The convolutional neural networks used were pretrained and fine-tuned, and they had an an encoder-decoder architecture. The ResNet22 encoder was pretrained with mammography images, while the ResNet34 encoder was pretrained with various color images. These encoders were paired with either a U-Net or a Feature Pyramid Network decoder. Additionally, U-Net model with random initialization was also tested. The five different models were trained and tested on the Oulu Dataset of Screening Mammography (9204 images) and on the Portuguese INbreast dataset (410 images) with two different loss functions, binary cross-entropy loss with soft Jaccard loss and a loss function based on focal Tversky index. The best models were trained on the Oulu Dataset of Screening Mammography with the focal Tversky loss. The best segmentation result achieved was a Dice similarity coefficient of 0.816 on correctly segmented masses and a classification accuracy of 88.7% on the INbreast dataset. On the Oulu Dataset of Screening Mammography, the best results were a Dice score of 0.763 and a classification accuracy of 83.3%. The results between the pretrained models were similar, and the pretrained models had better results than the non-pretrained models. In conclusion, deep transfer learning is very suitable for mammography mass segmentation and the choice of loss function had a large impact on the results.Rinnan massojen ...