Analysis of the response to neoadjuvant treatment in breast cancer using deep networks
Breast cancer is the most common type of cancer among women, it has the highest death rate among all types of cancer. Pre-surgical (neoadjuvant) treatment can improve the patient's prognosis, but it is not possible to predict whether the patient will respond to treatment. Previous studies were conducted to find characteristics capable of associating neoadjuvant treatment with patient response, some using conventional radiomics and others convolutional networks, mostly using private databases with few examples, those studies are non-reproducible. In this work, a deep learning model was proposed using images and clinical data from a public database (Duke Breast Cancer MRI), capable of extract characteristics of breast MRI images, and associate the attributes with prognosis. The 300 patients were divided between training/validation ($90\%$) and testing ($10\%$) using cross validation of three folds, where each fold has a different combination of patients, being $90\%$ for training and $10\%$ for validation. Using quantitative analysis of results generated from the trained model, it was concluded that the proposed model is capable of classifying patients that achieved complete response to neoadjuvant treatment. The results demonstrated superior accuracy when compared to the study by Cain et al. in the same database, mean AUC (area under the curve) from $0,70$ to $0,82$ and mean accuracy of $70\%$ for testing and $75\%$ for validation. The proposed model obtained competitive results compared to the literature in public databases, and it is not possible to fully reproduce the works of other authors. Considering conventional radiomics works, the experiments showed that the deep convolutional network method is also sensitive to the data. In this way, imaging methods can improve the results.