Advancing Perceptual Image Quality Assessment in Medical Imaging

Abstract

본 논문은 컴퓨터 단층 촬영(CT) 영상에서 객관적 이미지 품질 측정 기준(IQM)에 초점을 맞춘 이미지 품질 측정 기준을 개발하는 것을 목표로 한다. 이러한 IQM는 의료진의 의견을 진단 이미지 품질을 평가하기 위한 표준 점수로 간주하여 IQM과 의료진의 주관적 이미지 품질 인식 간의 상관관계를 통하여 탐구된다. 12개의 full-reference (FR) 및 no-reference (NR) IQM에 대한 포괄적인 비교는 다양한 노이즈 수준을 가진 CT 이미지 데이터 세트를 사용하여 수행되며 상관관계는 Pearson 선형 상관 계수(PLCC)와 Spearman의 순위 상관 계수(SROCC)를 사용하여 평가되었다. 그 결과 SSIM 및 PSNR과 같은 지표는 가장 높은 상관관계 점수를 보여주지만 의료진의 평가의 특정 측면을 완전히 포착하지는 못한다는 것을 증명하였다. 따라서 두 번째 연구는 NR 방법인 D2IQA(Deep Detector Image Quality Assessment) 개발을 통해 이러한 한계를 극복하고자 하였다. 이는 객체 감지 모델에 self-supervised learning을 사용하고 다른 NR-IQM 및 일부 FR-IQM과 비교하여 우수한 성능을 보여주었다. 마지막으로 세 번째 연구에서는 해당 모델의 계산 효율성의 한계를 해결하기 위해 EDIQA(Efficient Deep-Detector Image Quality Assessment)가 제안되었다. 이는 지식 증류와 task-based 방법을 사용하여 모델을 학습시키며, 그 결과 EDIQA는 다양한 영상 모달리티와 해부학적 영상에 걸쳐 우수한 일반화 기능을 보여주는 동시에 상당한 속도 향상을 달성하여 의료 영상에서 보편적인 IQM으로 사용될 수 있는 가능성을 보여주었다.;This paper aims to develop an image quality metric with a focus on objective image quality metrics (IQMs). These IQMs have been widely developed and utilized in the context of computed tomography (CT) imaging to optimize radiation doses. However, their correlation with a radiologist's subjective perception of image quality, which serves as the gold standard for assessing diagnostic image quality, remains largely unexplored. Furthermore, in the medical field, the assessment of image quality is based on how effectively an image provides the necessary information for physicians to make accurate diagnoses. It is crucial for the results of image quality assessment (IQA) to align with radiologists' opinions, as they are considered the gold standard in medical IQA. In pursuit of this objective, the initial study aims to investigate the correlation between subjective and objective quality metrics. A comprehensive comparison was conducted among twelve full-reference and no-reference IQMs, encompassing metrics such as root mean square error, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), multi-scale SSIM, gradient magnitude similarity deviation, feature similarity index, noise quality metric, visual information fidelity, natural image quality evaluator, blind/referenceless image spatial quality evaluator, perception-based image quality evaluator, and the model observer non-pre-whitening with eye filter (NPWE). The dataset employed in this investigation comprised CT images across seven distinct noise levels. The objective IQMs were employed to derive scores from this dataset, which were subsequently juxtaposed against the evaluations of three radiologists. The Pearson linear correlation coefficient (PLCC) and Spearman's rank order correlation coefficient (SROCC) were employed to quantify the extent of correlation. The findings revealed that SSIM and PSNR exhibited the highest correlation scores in terms of PLCC and SROCC, yet certain aspects of the radiologists' assessment were not fully captured by these metrics. The second presented study focuses on the development of a no-reference image quality assessment (NR-IQA) method, which is particularly valuable in clinical settings compared to full-reference IQA (FR-IQA). Our approach involves the utilization of a novel self-supervised training strategy for object detection models, specifically designed to detect virtually inserted objects with geometrically simple forms. By employing this strategy, we introduce a new NR-IQA method called Deep Detector IQA (D2IQA), which enables automated quantitative assessment of CT image quality. Through extensive experimental evaluations conducted on clinical CT images, we demonstrate the robustness of our D2IQA in accurately computing perceptual image quality across varying relative dose levels. Furthermore, when assessing the correlation between the evaluation results of IQA metrics and the quality scores provided by radiologists, our D2IQA exhibits advantage over other NR-IQA metrics and even shows competitive performance compared to FR-IQA metrics. The objective of the third study is to develop an efficient IQA algorithm suitable for real-time applications. To address the limitations of existing methods in capturing diagnostic quality and computational efficiency, we propose a novel algorithm called Efficient Deep-detector Image Quality Assessment (EDIQA). Our approach incorporates knowledge distillation within a two-stage training procedure, where a task-based IQA model, namely modified D2IQA, serves as the teacher model, and a specialized student model is designed for effective learning. This framework enables the student model to compute image scores based on a task-oriented approach, eliminating the need for complex signal insertion and multiple predictions. Consequently, our proposed method achieves a substantial speed improvement of over 1.6e+4 times compared to the teacher model. Additionally, we have devised a deep learning architecture that facilitates the student model in learning hierarchical multi-scale features of the image, ranging from low-level to high-level semantic features. Through extensive experiments, our model demonstrates excellent generalization capabilities across different modalities and anatomical parts, thereby representing a significant step towards a universal IQA metric in the field of medical imaging.