GPEC: Efficient Pre-LLM Gaussian Process Embedding Correction for Cardiac Video Caption Generation
This work introduces Gaussian Process Embedding Correction (GPEC), a modular method to improve video captioning accuracy for medical ultrasound by correcting visual representations before the language model processes them. GPEC learns a residual correction that moves the projected visual data closer to an annotation-guided target, significantly boosting caption quality and clinical attribute recovery without retraining the main model.