Neural Networks with Attribute Analysis Module

The incidence of thyroid nodules is rising annually, making the automatic diagnosis of these nodules increasingly critical. Deep neural networks offer the potential to automatically learn and classify thyroid nodular features. However, conventional neural networks often focus solely on classification categories, lacking interpretability in their predictions. To address this, we propose a Convolutional Processing Module-Attribute Analysis Module (CPM-AAM) framework consisting of Convolutional Processing Module (CPM) and Attribute Analysis Module (AAM). The CPM contains several convolutional layers, and the AAM includes residual blocks and fully connected layers. The AAM first extracts thyroid nodular attributes, such as composition, calcification, boundary and echogenicity, from thyroid nodules and then integrates these features to classify the nodules. We evaluated the performance of three standard classification networks (VGG16, GoogLeNet, and ResNet50), two existing methods and Transformer, assessing their performance with and without the AAM. Experimental results demonstrate that incorporating the AAM not only enhances classification performance but also provides detailed attribute classification, thereby improving the interpretability of the results.