Preoperative prediction of sentinel nodal status using mammography images

Sammanfattning: Breast cancer is the most common cancer among women in Sweden, accounting for approximately 30 % of the cancer cases. The overall prognosis is good, but worsens if the cancer metastasizes from the primary tumor. In order to exclude or confirm lymph node metastasis in clinically node negative breast cancer, axillary lymph nodes are examined by sentinel lymph node biopsy (SLNB). Up to 85 % of the patients have benign sentinel nodes, and do not benefit therapeutically from SLNB. This project was part of a goal to decrease unnecessary surgeries by preoperative prediction of sentinel nodal status. A cohort of 800 patients, diagnosed with primary breast cancer in Scania, Sweden, between 2009 and 2012 was studied. The cohort was previously used by Dihge et al. to predict axillary lymph node metastasizing, using a multilayer perceptron (MLP) on clinicopathological data. The aim of this project was to determine whether including information from mammograms would improve the artificial neural network’s prediction. A similar MLP to that of Dihge et al. was constructed, and convolutional neural networks were used to extract features from the mammography images (n = 705). The features were used as additional input to the MLP. The results were evaluated with area under the ROC curve (AUC) score. The addition of features from mammograms did not improve the predictions. The MLP’s AUC-score without features from mammograms was 0.7190 (std 0.0465), it decreased to 0.6573 (std 0.0470) when features from mammography images were added. Nevertheless, the results have demonstrated behaviors of the models and may therefore be used to guide future attempts at using mammograms to improve sentinel lymph node prediction.