A Transformer-Enhanced Generative AI Framework for Lung Tumor Segmentation and Prognosis Prediction

Abstract

Lung cancer remains a leading cause of cancer-related mortality worldwide, with its effective management heavily reliant on the precise segmentation of tumors from medical imaging and accurate prognostic stratification. While convolutional neural networks (CNNs) have become the de facto standard for medical image segmentation, their inherent locality limits their ability to model long-range contextual dependencies, which are crucial for accurate delineation of complex tumor boundaries and heterogeneous textures. Furthermore, the integration of segmented tumor characteristics into a robust prognostic model presents a significant challenge. This paper proposes a novel Transformer-Enhanced Generative AI Framework that synergistically combines a hybrid CNN-Transformer architecture for segmentation with a generative adversarial network (GAN) for data augmentation and a recurrent prognostic module. The segmentation module leverages the Transformer's self-attention mechanism to capture global contextual information, enhancing the delineation of tumor margins. The generated high-quality segmentations are then utilized to extract rich radiomic features, which feed into a predictive model for patient survival analysis. Experimental results on the NSCLC-Radiomics dataset demonstrate that our framework achieves a Dice similarity coefficient of 0.891, outperforming several state-of-the-art baselines. The prognostic model achieves a concordance index of 0.741, indicating a strong predictive power for survival outcomes. This integrated approach demonstrates the significant potential of transformer-enhanced generative models in advancing precision oncology for lung cancer care.