| dc.description.abstract | CONTEXT: Cancer is nowadays one of the leading public health problems worldwide and
breast cancer is one of the most common in women. The prognosis and overall patient survival significantly decrease when breast cancer metastasizes. The evaluation of the presence of metastatic cells in the sentinel lymph node is currently the gold standard for the diagnosis of metastases, but the examination process is time-consuming for the pathologist and susceptible to failures, especially for the detection of micrometastasis. Advances in the histopathological image digitalization and deep learning added a highlight for the study of models that can assist in tasks related to micrometastases diagnosis in breast cancer. OBJECTIVE: Thus, this research aims to assist in this investigation, through the investigation of a deep learning model capable of detecting breast cancer micrometastases at an efficiency comparable to pathologists. METHODOLOGY: To achieve this objective, our architecture is divided into two main tasks. The first consists of a convolutional neural network to perform a patch-level classification at the level of fragments of the original image with full resolution - which in this work we will refer as a patch -. Afterwards, we will perform the second task which is responsible for the
segmentation task at the pixel level to extract the metastatic areas of the images and measure it, in order to identify the micrometastases. For such training, we are using the Camelyon16 challenge dataset. Therefore, the evaluation metrics of our model are based on the baselines evaluated by this challenge. RESULTS: For the partial results, our classification task achieved AUC = 0.998, in the isolated tests carried out at the fragmented level of the slide, resulting in a F1Score = 1.00 for the negative class and F1Score = 0.99 for positive class, not generating false negatives in the partial steps. Our segmentation task has reached the result of IoU − Score : 0.5434 F I − Score : 0.64818. The final results were found through the reconstruction of the segmented images. Although we obtained good results in the partial and isolated tests for each task for the slides fragments, they did not corroborate with the final results of the slide produced at the end of the framework, thus those results did not demonstrate the metrics found in the partial tests, not being possible to locate the regions of metastasis precisely. However, there is still room for improvements in the model, and the experimental results indicate that the method can contribute to the proposed study. As we chose to work with the images in the highest resolution, dividing them into patches and having a two-layer neural network model - classification and segmentation, the processing time of a single slide by the proposed framework is up to 2 hours. CONCLUSION: Our results indicate
that, with the improvement of implementation, this model has the potential to meet the
proposed contributions and this master thesis indicates new directions regarding new analyzes and tests to implement improvements on it. | en |
