The exponential growth of deep learning networks has allowed us to tackle complex tasks, even in fields as complicated as medicine. However, using these models requires a large corpus of data for the networks to be highly generalizable and with high performance. In this sense, data augmentation methods are widely used strategies to train networks with small data sets, being vital in medicine due to the limited access to data. A clear example of this is magnetic resonance imaging in pathology scans associated with cancer. In this vein, we compare the effect of several conventional data augmentation schemes on the ResNet50 network for brain tumor detection. In addition, we included our strategy based on principal component analysis. The training was performed with the network trained from zeros and transfer-learning, obtained from the ImageNet dataset. The investigation allowed us to achieve an F1 detection score of 92.34%. The score was achieved with the ResNet50 network through the proposed method and implementing the learning transfer. In addition, it was also concluded that the proposed method is different from the other conventional methods with a significance level of 0.05 through the Kruskal Wallis test statistic.
Brain tumors are usually fatal diseases with low life expectancies due to the organs they affect, even if the tumors are benign. Diagnosis and treatment of these tumors are challenging tasks, even for experienced physicians and experts, due to the heterogeneity of tumor cells. In recent years, advances in deep learning (DL) methods have been integrated to aid in the diagnosis, detection, and segmentation of brain neoplasms. However, segmentation is a computationally expensive process, typically based on convolutional neural networks (CNNs) in the UNet framework. While UNet has shown promising results, new models and developments can be incorporated into the conventional architecture to improve performance. In this research, we propose three new, computationally inexpensive, segmentation networks inspired by Transformers. These networks are designed in a 4-stage deep encoder-decoder structure and implement our new cross-attention model, along with separable convolution layers, to avoid the loss of dimensionality of the activation maps and reduce the computational cost of the models while maintaining high segmentation performance. The new attention model is integrated in different configurations by modifying the transition layers, encoder, and decoder blocks. The proposed networks are evaluated against the classical UNet network, showing that our networks have differences of up to an order of magnitude in the number of training parameters. Additionally, one of the models outperforms UNet, achieving training in significantly less time and with a Dice Similarity Coefficient (DSC) of up to 94%, ensuring high effectiveness in brain tumor segmentation.
La resonancia magnética funcional en estado de reposo (rs-fMRI) es una de las técnicas más relevantes en exploración cerebral. No obstante, la misma es susceptible a muchos factores externos que pueden ocluir la señal de interés. En este orden de ideas, las imágenes rs-fMRI han sido estudiadas desde diferentes enfoques, existiendo un especial interés en las técnicas de eliminación de artefactos a través del Análisis de Componentes Independientes (ICA por sus siglas en inglés). El enfoque es una herramienta poderosa para la separación ciega de fuentes donde es posible eliminar los elementos asociados a ruido. Sin embargo, dicha eliminación está sujeta a la identificación o clasificación de las componentes entregadas por ICA. En ese sentido, esta investigación se centró en encontrar una estrategia alternativa para la clasificación de las componentes independientes. El problema se abordó en dos etapas. En la primera de ellas, se redujeron las componentes (volúmenes 3D) a imágenes mediante el Análisis de Componentes Principales (PCA por sus siglas en inglés) y con la obtención de los planos medios. Los métodos lograron una reducción de hasta dos órdenes de magnitud en peso de los datos y, además, demostraron conservar las características espaciales de las componentes independientes. En la segunda etapa, se usaron las reducciones para entrenar seis modelos de redes neuronales convolucionales. Las redes analizadas alcanzaron precisiones alrededor de 98 % en la clasificación e incluso se encontró una red con una precisión del 98.82 %, lo cual refleja la alta capacidad de discriminación de las redes neuronales convolucionales.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.