| Abstrakt: |
Learning discriminate representations from images plays crucial role in medical image analysis. The attention mechanism, on the other hand, leads to breakthrough results in the computer vision field by allowing models to provide varying levels of focus across image regions. In this work, we present Tumor Aware Net, an end-to-end trainable attention based Convolutional Neural Network that learns effective representations from Magnetic Resonance (MR) images suitable for effective tumor recognition. The proposed model employs a Sparse Convolutional Denoising Autoencoder (SCDA) to project the higher dimensional MR image representations to a lower dimensional space with improved discrimination. These lower dimensional descriptors are passed through an attention module, which prioritizes tumor descriptors over the rest. Furthermore, the proposed SCDA is trained jointly with the Neural induced Support Vector Classifier (NSVC) to achieve maximum margin separation. The proposed model has been validated on several publicly available benchmark datasets for tumor recognition. Based on the outcomes of the experimental studies, we claim that the proposed model favours stability and complements the learned representations when combined with attention. Despite its simplicity in terms of model parameters, the proposed model outperforms existing models for tumor type categorization. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink