Translational Medical Intelligence

We develop deep learning-based approaches to recover high-quality images from noisy measurements or degraded images. Our goal is to enhance image quality while preserving the underlying textures and structures by adapting recent computer vision techniques to the specific tasks. Building on this work, we aim to extend our research to enable more reliable image interpretation and provide feasible solutions to real-world problems. We conduct fundamental research to determine how reliable AI model is in bridging the gap between its prediction and ground truth. We assess model confidence based on uncertainty estimation and control the use of learned information and the measurement knowledge. Our goal is to establish AI model that can maintain meaningful information without hallucination. We develop AI methods for understanding dental images and connecting visual findings with clinical information. Our research spans lesion detection, anatomically guided lesion synthesis, data augmentation, and report-guided annotation using panoramic radiographs and cone-beam CT. Building on this work, we are exploring multimodal models and 3D anatomical reconstruction to support dental image interpretation. We design highly versatile new CT imaging systems by jointly implementing complex acquisition geometry, calibration, and image reconstruction. We are studying efficient reconstruction algorithms that can increase scanning coverage, even within the constraints of limited spaces. Our goal is to develop fast, memory-efficient methods for prototype imaging system applicable to real industrial CT. We investigate more clinically meaningful evaluation metrics in real practice to overcome the methodological representability limitations of widely used metrics. We pursue to design an image quality