TTT-SEG: a multi-structure segmentation on cardiac MRI using vision testing-time training network

We propose TTT-SEG, a Test-Time Training Segmentation model, for automatic cardiac organ segmentation on MRIs. Manual contouring in MR imaging is time-consuming and labor-intensive, making automatic segmentation crucial for efficient therapy workflows. Despite advancements in Convolutional Neural Networks (CNNs) and Vision Transformer-based Networks achieving state-of-the-art accuracy, their performance is hindered by the lack of large-scale MRI datasets, leading to poor generalization on new patients. TTT-SEG addresses this issue by integrating a CNN with a novel learning-to-learn strategy. Our model features a dual-loop mechanism: a Reconstruction Learning Loop (RLL) for learning self-supervised features and an Adaptive Learning Loop (ALL) for segmentation features. Unlike the traditional networks, the RLL learns even during testing, improving segmentation accuracy for specific samples. Evaluating TTT-SEG on the Automated Cardiac Diagnosis Challenge (ACDC) dataset shows superior performance in segmenting the left and right ventricles and myocardium, outperforming methods like Vnet, nnUnet, SwinUNETR, and nnFormer. These results demonstrate TTT-SEG's potential to enhance cardiac organ segmentation and improve cardiac-related diagnosis and therapy workflows.