Fast and scalable annotation-free LV-centered ROI localisation in stress perfusion cardiac MRI.

Fast and scalable localisation of a left ventricle (LV)-centered region of interest (ROI) is essential for cardiac image analysis, particularly in stress perfusion cardiovascular magnetic resonance (CMR), where manual annotation is time-consuming and labelled datasets are scarce. This study proposes a fully automated, annotation-free pipeline that integrates Fast Fourier Transform-based localisation, Sobel edge refinement, and a fallback heuristic to generate consistent LV-centered ROIs directly from raw perfusion frames, without requiring task-specific annotations. A circular ROI is applied as a post-processing step to ensure consistent coverage of the ventricular cavity and surrounding myocardium while preserving diagnostically relevant boundary information. To assess whether learning-based models can replicate these pseudo-annotations without perfusion-specific training, a clinician-verified subset of 295 FFT-Sobel annotations was used to benchmark a pretrained U-Net and a one-shot similarity model. The FFT + Sobel method achieved successful ROI localisation in 81.5% of 460 frames at 0.002 s per frame. The U-Net achieved a Dice score of 88.6% and IoU of 80.1%, but only 72% full ROI detection, while the one-shot model showed limited spatial agreement (Dice 26.7%, IoU 16.8%). These findings demonstrate that the proposed classical pipeline provides a fast, robust, and scalable solution for LV-centered ROI localisation in annotation-free perfusion CMR, offering a consistent and clinically meaningful representation for downstream cardiac image analysis.