Click here to


Are you sure ?

Yes, do it No, cancel

BEST IN PHYSICS (IMAGING): A Deep Learning Based Model Observer for Low-Contrast Object Detection Task in X-Ray Computed Tomography

H Gong, L Yu , S Leng , C McCollough* , Mayo Clinic, Rochester, MN


(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM

Room: Room 202

Purpose: Model observers (MO) have become popular for quantitative and objective quality assessment in X-ray computed tomography (CT), but are usually limited to phantom data. This study aims to develop a deep-learning-based MO that operates on patient images and is well-correlated with human observer (HO) results in a realistic low-contrast lesion detection task that involves real anatomical background and multi-slice scrolling image review.

Methods: The MO was developed using transfer learning to integrate a pre-trained deep-convolutional-neural-network (DCNN) with partial-least-square regression (PLSR). Assuming similarity between DCNN and the human visual system, early layers of the DCNN were used as a deep feature extractor. PLSR was trained over deep features to generate MO test statistics. Seven abdominal CT exams were retrospectively collected from the same CT system. CT images of a real liver lesion were modified to generate lesion models with four lesion sizes (5, 7, 9, & 11 mm) and three contrast levels (15, 20 & 25 HU). These were inserted into patient liver images using a projection-based method. Noise was added to projection data to synthesize CT exams with half and quarter of routine dose levels. Four medical physicists performed a two-alternative-forced-choice (2AFC) detection task (with multi-slice scrolling) in patient images from 12 experimental conditions. The MO performed a 2AFC task on the same data. An internal noise component was added to the MO test statistics to calibrate to HO performance at one experimental condition.

Results: MO results were well correlated with HO for the 2AFC detection task with patient liver images. The corresponding Pearson correlation coefficient was 0.968 (95% confidence interval [0.888, 0.991]). Bland-Altman agreement analysis did not indicate statistically significant differences.

Conclusion: This study showed the potential to use the proposed deep-learning-based MO to directly assess diagnostic quality of patient images for realistic, clinically-relevant CT detection tasks.


Image Analysis, Image Processing


IM- CT: Quantitative imaging/analysis

Contact Email