Twelve Fused Models by Fusing Four Types of Transformer-Based Tuners on Three Base UNet Architectures for Carotid...

Background/Objectives: UNet-based models dominate medical image segmentation. Transformers have been added as an internal variant to these UNet-based architectures to improve feature learning. However, they have limitations in generalization and computational efficiency. Motivated by this idea, we have designed a two-stage novel hybrid segmentation framework, where tuners are added in cascade to the base architectures. Methods: Three sets of base UNets were designed, namely: B1: UNet1p, B2: UNet2p, and B3:UNet3p, and four sets of transformer-based tuners were designed, namely: T1:Transformer-augmented UNet, T2: Attention-guided UNet, T3: Swin Transformer-based UNet, and T4: Pyramid-based network, leading to 12 fused systems that combine three base UNets and four Tuners, namely: F1: B1 + T1, F2: B1 + T2, F3: B1 + T3, F4: B1 + T4; F5: B2 + T1, F6: B2 + T2, F7: B2 + T3, F8: B2 + T4, F9: B3 + T1, F10: B3 + T2, F11: B3 + T3, F12: B3 + T4. Results: The two-hybrid segmentation models are more effective and reliable than the single-stage UNet architecture. B3 + T4 achieved a Dice of 94.14% and Jaccard of 88.7%, surpassing prior baselines by 4.2% and 6.8%. It reduced cIMTE to 0.014 mm, a 36% improvement and the lowest reported to date, with cLIE and cMAE errors lowered by 40%. Conclusions: All 12 hybrid automated transformer-based models are highly accurate and reliable for wall segmentation in carotid ultrasound; they are a powerful paradigm for cardiovascular risk.