Abstract
Accurate prediction of the tumor’s future imaging features can provide its complete growth evolution and more detailed clinical parameters. The existing longitudinal models tend to lose detailed growth information and make it difficult to model the complete tumor development process. In this paper, we propose the Static–Dynamic coordinated Transformer for Tumor Longitudinal Growth Prediction (SDC-Transformer). To extract the static high-level features of tumors in each period, and to further explore the dynamic growth associations and expansion trend of tumors between different periods. Aiming at the insensitivity to local pixel information of the Transformer, we propose the Local Adaptive Transformer Module to facilitate a strongly coupled status of feature images, which ensures the characterization of tumor complex growth trends. Faced with the dynamic changes brought about by tumor growth, we introduce the Dynamic Growth Estimation Module to predict the future growth trend of the tumor. As a core part of SDC-Transformer, we design the Enhanced Deformable Convolution to enrich the sampling space of tumor growth pixels. And a novel Cascade Self-Attention is performed under multi-growth imaging to obtain dynamic growth relationships between periods and use dual cascade operations to predict the tumor’s future expansion trajectories and growth contours. Our SDC-Transformer is rigorously trained and tested on longitudinal tumor data composed of the National Lung Screening Trial (NLST) and collaborative Shanxi Provincial People’s Hospital. The RMSE, Dice, Recall, and Specificity of the longitudinal prediction results reach 11.32, 89.31%, 90.57%, and 89.64%, respectively. This result shows that our proposed SDC-Transformer model can achieve accurate longitudinal prediction of tumors, which will help physicians to establish specific treatment plans and accurately diagnose lung cancer. The code will be released soon.
•We apply the improved Transformer approach to longitudinal tumor growth studies, using its excellent remote modeling ability to explore tumor growth trends over long time series.•We design the Local Adaptive Transformer Module, which effectively alleviates the Transformer’s insensitivity to local pixels.•We propose the Enhanced Deformable Convolution Module enriches the sampling space for feature images and learns the pixel locations of future tumor growth.•A novel cascade self-attention is adopted to establish a long-range dependence between tumor growth pixels.