Fully Convolutional Transformer for Malignancy Estimation of Pulmonary Nodule

The fine-grained characteristics inherent in medical imaging represent one of the most significant distinctions from regular images, a distinction that has resulted in suboptimal performance of Transformers in this domain. Traditional convolutional neural networks, exemplified by ResNet, leverage residual structures to deepen the network, thereby offering hierarchical feature information for fine-grained attribute classification, a capability not present in existing Transformer-based models. In this study, we utilize a fully convolutional Transformer model that merges broad convolution with the self-attention mechanism. This model's overall architecture adheres to the traditional down-sampling pattern found in ResNet, effectively modeling both long and short-term dependencies in the input. The model operates in two stages: extracting long-term semantic information and capturing layered global attributes from the features, resulting in a more compact and robust structure overall. We trained this model using samples and labels extracted from the LIDC-IDRI dataset, enabling us to perform a detailed classification of the malignancy levels of lung nodules. According to our results, this model effectively mitigates overfitting, outperforming both Transformer and ResNet on this dataset. Furthermore, it demonstrates a high level of predictive accuracy on a test set with a limited number of small cell lung cancer samples.