MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Comparison of Noise Analysis and Deep Learning-Based Image Quality Assessment (IQA) Methods for Thoracic Computed Tomography (CT)

B Grant1*, J Lee2 , J Chung3 , I Reiser2 , L Lan2 , J Papaioannou3 , M Giger2 , (1) Western Kentucky University, Bowling Green, KY, (2) The University of Chicago, Chicago, IL, (3) University of Chicago Medicine, Chicago, IL,

Presentations

(Tuesday, 7/31/2018) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 1

Purpose: Since CT uses ionizing radiation, image quality should be optimized to produce diagnostic quality while minimizing radiation dose to the patient. Diagnostic quality of an image is a complex parameter that not only depends on image noise but on other contextual parameters such as the patient's anatomy and the pathology of interest. Therefore, it is important to choose the appropriate image quality assessment (IQA) technique. The purpose of this work is to investigate and compare a conventional noise analysis method and two deep learning-based methods.

Methods: Image quality analysis was conducted on a data set of 74 cases of CT scans for interstitial lung disease (ILD). Each case contained an average of 300 high-resolution slices resulting in a total of 21,257 slices. The cases were rated by a board-certified radiologist as diagnostic, but low quality or diagnostic and high quality. Noise analysis was performed using the global noise level, which is found as the mode of local standard deviation of pixel values surrounding each image pixel. Classification of features extracted by a pre-trained convolutional neural network using a support vector machine as well as fine-tuning of a pre-trained network were assessed as methods of deep-learning IQA. Area under the receiver operating characteristic curve (AUC) was used as the performance metric.

Results: The deterministic global noise level analysis achieved an AUC of 0.58 (SE: 0.005). Feature extraction method performed similarly achieving an AUC of 0.63 (SE: 0.02). Fine-tuning method performed significantly better achieving an AUC of 0.78 (SE: 0.01).

Conclusion: While this work is a preliminary study with a relatively small number of cases, within this data set, deep learning was better able to classify image quality than the traditional method, demonstrating the potential of deep learning as an IQA technique.

Funding Support, Disclosures, and Conflict of Interest: J. Chung: Consultant, speakers bureau, Genentech, Inc.; Consultant, speakers bureau; Boeringer Ingelheim, Inc.; Consultant, Veracyte, Consultant, ACI Clinical M. Giger: Stockholder: Hologic, Inc., Quantitative Insights, Inc.; Shareholder, QView Company; Co-founder, Quantitative Insights, Inc.; Royalties: Hologic, Inc., General Electric Company, MEDIAN Technologies, Riverain Technologies, LLC, Mitsubishi Corporation, Toshiba Medical Systems Corporation

Keywords

Image Analysis, Computer Vision, Lung

Taxonomy

IM/TH- Image Analysis (Single modality or Multi-modality): Machine learning

Contact Email