Lung tumor growth prediction of follow-up via Spatio-Temporal Convolutional Transformer

Precise forecasting of lung tumor growth is essential for devising effective treatment strategies and managing early-stage lung cancer. However, existing methods lack intuitive judgments for tumor growth and evolution patterns, as well as clinical information. To address this problem, this paper introduces a versatile and easy-to-train architecture called the Spatio-Temporal Convolutional Transformer (ST-ConvTransformer) to facilitate the prediction of early-stage tumor growth. The ST-ConvTransformer is composed of two transformer submodules that utilize the same self-attention layers. During the extraction of visual features from medical images, textual features from clinical information are also extracted and integrated. Simultaneously, the visual encoder submodule incorporates the temporal and spatial information of tumors. Extensive experiments are conducted on a dataset with 2800 clinical subjects. The proposed model achieves a Precision of 80.87%, Recall of 89.74%, and Dice Similarity Coefficient of 84.24%. These results demonstrate the potential of the ST-ConvTransformer as a reliable clinical-aided tool for predicting lung tumor growth.