| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 423.07 KB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Os sistemas CAD auxiliam a deteção e diferenciação de lesões benignas e malignas, aumentando a performance no diagnóstico do cancro da mama. Uma vez que as lesões da mama estão fortemente correlacionadas com a forma e contorno, neste estudo, foram aplicados dois métodos diferentes para classificação de lesões em imagens de mamografia. O primeiro consiste em medidas quantitativas baseadas na dimensão fractal, calculadas através da aplicação do método “boxcounting”, diretamente em imagens de lesões segmentadas antes e após a aplicação de um algoritmo de dilatação/erosão. O segundo método baseou-se na aplicação de Redes Neuronais de Convolução (CNN), as quais têm demonstrado um elevado grau de sucesso, na deteção e classificação de patologias em diferentes modalidades de imagem médica, incluindo a mamografia. De forma a ultrapassar a limitação do reduzido número de amostras disponíveis nas bases de dados de mamografia, foi aplicado o método de “transfer learning”, no qual três modelos CNN prétreinados num grande conjunto de dados foram ajustados de forma a permitir a classificação de lesões em imagens de mamografia antes e após a aplicação de um processo de normalização. Também foi avaliada a performance de uma SVM com a utilização de características extraídas das CNN isoladamente ou combinadas com “handcrafted features”. Os resultados obtidos são encorajadores e demonstram que tanto o uso da medida de dimensão fractal como das CNN pode ajudar a melhorar o diagnóstico automático de lesões em mamografia, o que se traduziu em valores de AUC até 81,3%.
ABSTRACT - Computer‑Aided Systems can improve the performance of breast cancer diagnosis, helping the differentiation and classification of breast benign and malignant lesions. Breast lesions are strongly correlated with their shape, therefore, in this work, firstly, were used quantitative measures based on fractal dimension, which can help characterize the smoothness or the roughness of mammographic lesion shape, being calculated through the box‑counting method, directly from manually segmented lesions, and after applying a region growing/erosion algorithm. Then, Convolutional neural networks (CNN), which recently have been successfully used in the medical field to detect and classify pathologies in different imaging modalities, were applied to mammographic images. Since the mammographic databases have a restricted number of samples, one way to solve this problem is using a transfer learning approach, in which a CNN, previously pre-trained with a large amount of labeled non-medical data, is subsequently fine-tuned using a smaller dataset of mammographic images. In this study were applied three different pre-trained networks were made an evaluation if their performance to classify lesions in mammograms is affected by the use, or not, of normalized images. Also evaluated was the performance of a support vector machine fed with features extracted from the CNN and the combined use of handcrafted features to complement the CNN-extracted features. The obtained results are encouraging and reveal that both the use of fractal dimension and CNN can help to improve computer-aided diagnosis of mammographic lesions, with AUC values around 81.3%.
ABSTRACT - Computer‑Aided Systems can improve the performance of breast cancer diagnosis, helping the differentiation and classification of breast benign and malignant lesions. Breast lesions are strongly correlated with their shape, therefore, in this work, firstly, were used quantitative measures based on fractal dimension, which can help characterize the smoothness or the roughness of mammographic lesion shape, being calculated through the box‑counting method, directly from manually segmented lesions, and after applying a region growing/erosion algorithm. Then, Convolutional neural networks (CNN), which recently have been successfully used in the medical field to detect and classify pathologies in different imaging modalities, were applied to mammographic images. Since the mammographic databases have a restricted number of samples, one way to solve this problem is using a transfer learning approach, in which a CNN, previously pre-trained with a large amount of labeled non-medical data, is subsequently fine-tuned using a smaller dataset of mammographic images. In this study were applied three different pre-trained networks were made an evaluation if their performance to classify lesions in mammograms is affected by the use, or not, of normalized images. Also evaluated was the performance of a support vector machine fed with features extracted from the CNN and the combined use of handcrafted features to complement the CNN-extracted features. The obtained results are encouraging and reveal that both the use of fractal dimension and CNN can help to improve computer-aided diagnosis of mammographic lesions, with AUC values around 81.3%.
Description
Keywords
Manografia CAD Dimensão fractal CNN Transfer learning Deep learning Mammography Fractal dimension
Citation
Perre AC, Freire LC. Avaliação da caracterização de lesões em mamografia com recurso a sistemas CAD. Roentgen. 2021;2(2):42-9.
Publisher
Associação NUCLIRAD