H Gong*, N Huber, C Koo, T Johnson, A Inuoe, J Marsh, J Thorne, S Leng, J Fletcher, C McCollough, L Yu, Mayo Clinic, Rochester, MN

• 
  H Gong

• 
  

TH-D-TRACK 3-6 (Thursday, 7/29/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Purpose: Deep convolutional neural network (CNN)-based denoising techniques have been used in clinical CT to improve image quality and reduce radiation dose. Objective assessment of its performance is desired but has been challenging due to its highly non-linear characteristics. We previously developed a deep-learning model-observer (DLMO) that can be directly applied to patient CT images. The purpose of this work is to validate the DLMO in a lung-nodule detection task in real patient images denoised by a CNN-based technique.

Methods: Fifty nodule-free patient chest CT exams were retrospectively collected. Nodule-present cases at lower radiation dose levels were synthesized using clinically-validated projection-domain lesion- / noise-insertion tools: 10 experimental conditions involving two nodule types (ground-glass-nodule, partial-solid-nodule), three nodule sizes (3.4–7.4 mm) and four radiation dose levels (10%, 25%, 50%, and 100% of routine-dose, quality-reference-mAs 50). A ResNet-based CNN-denoising model was trained using only nodule-free training cases at 25% routine-dose, and deployed to independent testing cases across four dose levels. DLMO used sliding window to analyze local features and generate pixel-wise test-statistics, and used an internal-noise-component to calibrate the performance to human level. DLMO was established with two
methods: DLMO-1 (baseline method) – DLMOs were separately trained and deployed at each dose level; DLMO-2 (more computationally-efficient method) – one DLMO was trained at one dose level and deployed to other levels. Two fellowship-trained thoracic radiologists and one general radiologist interpreted the same DLMO test cases. Area-under-curve of localization-ROC was used as figure-of-merit.

Results: In preliminary study, DLMO-1 and DLMO-2 performance yielded strong correlation with the averaged human performance (Pearson’s correlation coefficient: DLMO-1 – 0.995, P-value 0.0049; DLMO-2 – 0.996, P-value 0.004). Mean performance difference was small (DLMO-1 – 1.02%; DLMO-2 – 1.47%).

Conclusion: DLMO demonstrated the potential for efficient and objective image quality assessment in realistic lung-nodule detection tasks with CNN-based CT denoising.

Funding Support, Disclosures, and Conflict of Interest: Research reported in this work was supported by the National Institutes of Health under award number R01 EB017095. Research support is provided to Mayo Clinic from Siemens Healthcare GmbH, unrelated to this work.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email