SUN-811 Adrenal incidentaloma segmentation via efficient fully convolutional transformers

Abstract Disclosure: J. Park: None. J. Heo: None. J. Maeng: None. S. Park: None. H. Jeon: None. Y. Hwang: None. D. Lee: None. Background: Adrenal incidentaloma, a tumor identified incidentally through imaging, poses challenges due to its asymptomatic presentation and the need for precise identification amidst high-resolution imaging data. This study introduces the Efficient Fully Convolutional Transformer (eFCT), a novel deep learning model designed to achieve high segmentation accuracy for adrenal incidentaloma while significantly reducing computational demands. Methods: The eFCT model incorporates two key innovations: (1) Non-attention modules, which reduce the spatial resolution of input features at high-resolution layers, thereby minimizing computational overhead; and (2) Vast-focus modules, which expand the receptive field using non-consecutive dilation ratios to enhance the extraction of global contextual features without increasing computational cost. The model was trained and validated on a curated dataset of 33,003 2D abdominal CT images from 264 patients, annotated with segmentation masks identifying adrenal incidentalomas. Training utilized binary cross-entropy and Dice loss, with deep supervision applied to intermediate layers to improve learning stability. Performance metrics included Dice score, multiply-accumulate operations (MACs), and throughput (images per second), benchmarked against the Fully Convolutional Transformer (FCT) baseline. Results: The eFCT model demonstrated a Dice score of 0.870, surpassing the baseline (0.795) by 0.075 points. Computational efficiency was markedly enhanced, with a reduction in MACs by 11.3× and a 27.8× increase in inference speed compared to the baseline. Ablation studies confirmed the effectiveness of the non-attention modules and vast-focus design in balancing accuracy and efficiency. Qualitative analyses highlighted the eFCT’s superior ability to segment small, complex tumor regions and distinguish true positives from visually similar structures, reducing false negatives and false positives. Conclusion: The eFCT model represents a breakthrough in medical image segmentation, particularly for high-resolution tasks with constrained computational resources. Its lightweight architecture and enhanced accuracy make it a viable solution for clinical workflows, potentially extending to other medical imaging challenges. Future directions include adapting the model for multi-class segmentation tasks and evaluating its performance across diverse datasets. Presentation: Sunday, July 13, 2025