Efficient Deep-Detector Image Quality Assessment Based on Knowledge Distillation

Abstract

An efficient deep-detector image quality assessment (EDIQA) is proposed to address the need for an objective and efficient medical image quality assessment (IQA) without requiring reference images or ground-truth scores from expert radiologists. Existing methods encounter limitations in meeting diagnostic quality and computation efficiency, especially when reference images are unavailable. The proposed EDIQA leverages knowledge distillation in a two-stage training procedure, using a task-based IQA model and the modified deep-detector IQA (mD2IQA) as the teacher model and novel student model designed for effective learning. This approach enables the student model to compute image scores based on a task-based approach without complex signal insertion and multiple predictions, resulting in a speed improvement of over 1.6e+4 times compared to the teacher model. A deep-learning architecture is developed to allow the student model to learn hierarchical multiscale features of the image from low- to high-level semantic features. Rigorous evaluations demonstrate the generalizability of the proposed model across various modalities and anatomical parts, indicating a step toward a universal IQA metric in medical imaging. © 1963-2012 IEEE.