Ultra-High-Resolution Photon-Counting-Detector CT with a Dedicated Denoising Convolutional Neural Network for Enhanced Temporal Bone Imaging

Data Collection

This retrospective study was approved by our institutional review board and was Health Insurance Portability and Accountability Act–compliant, with informed consent waived. The methodology proposed in the Checklist for Artificial Intelligence in Medical Imaging (CLAIM; https://pubs.rsna.org/page/ai/claim) (Supplemental Data) was followed. Data from 6 adult clinical temporal bone PCD-CT scans were used for CNN training and validation, while 20 independent patient cases were used for testing. All examinations were conducted in UHR mode (120 × 0.2 mm collimation, 120 kV) on a dual-source PCD-CT (NAEOTOM Alpha; Siemens), with 1-second rotation time, 0.5 helical pitch, and automatic exposure control with 220 CARE (Siemens) kiloelectron volt Image Quality (IQ) level, resulting in a volume CT dose index of 34 mGy for standard size patients. The training and validation data set included 1885 CT images, reconstructed using both thin-slice (0.2-mm slice thickness with an 0.1-mm increment) and thick-slice (0.4-mm slice thickness with an 0.2-mm increment) settings. Quantum iterative reconstruction was applied at strength 1 (QIR1) for thin slices and strength 3 (QIR3) for both thin and thick slices. All images used the sharpest available head-regular reconstruction kernel (Hr96) and were processed with a 1024 × 1024 matrix size within an 80-mm clinical standard field of view. The trained CNN was then applied to test cases, using images reconstructed at 0.2-mm thickness with the Hr96 kernel and QIR1. For reference, images reconstructed with the clinical routine kernel (Hr84) at a 0.2-mm slice thickness and QIR3 were also collected.

Dedicated Denoising CNN Training Workflow

Figure 2 outlines our denoising CNN training workflow that begins by creating “Noise thin” images. These are generated by subtracting thin-slice iterative reconstructions with strengths 1_(QIR1_thin) and 3 (QIR3_thin), both set at a 0.2-mm slice thickness and increment, at the same anatomic location (higher strength settings perform more aggressive denoising). To prevent overfitting and introduce variability, we applied spatial decoupling techniques to Noise thin images through random translations (ranging from 1 to 16 pixels) and inversions (using multipliers of +1 or −1) to create a set of randomized noise images.

• Download figure
• Open in new tab
• Download powerpoint

FIG 2.

The overall workflow of the proposed deep CNN denoising method. All training data originated from patient image series reconstructed using 2 iterative reconstruction strengths, QIR1 and QIR3, with thin-slice (0.2-mm) and thick-slice (0.4-mm) thicknesses, respectively. A multiple-slice input strategy was implemented to enhance the performance of the CNN.

Additionally, a set of thicker-slice images (0.4-mm thickness, 0.2-mm increment), referred to as “QIR3_thick,” was reconstructed as a low-noise reference. CNN inputs were formed by combining noise-only and reference patches (QIR3_thick + α × Noise_thin) from 7 adjacent slices, with α empirically set at 2.0 to balance noise reduction and detail preservation.¹⁵ The central slice patch of QIR3_thick was the training target. As demonstrated in a previous study,¹⁴ training the CNN with thick reference images results in significantly improved noise reduction compared with using single-slice reference images. Moreover, the performance of the CNN remains consistent when applied to single-slice images, even though it was trained on thick images. This outcome is because the network focuses on learning noise patterns rather than the underlying tissue structures. Finally, the trained denoising efficacy of the CNN was then tested on thin QIR1 images.

Network Architecture and Training Details

We used a simplified U-Net architecture¹⁷ with 9 modules for our study. Each module sequentially performs convolution, batch normalization, and exponential linear unit activation operations. The architecture includes max pooling layers, convolution transpose operators, and concatenation to maintain input-output similarity. The mean-squared-error loss function was optimized during training. Our final training set comprised 18,864 patches of 128 × 128 pixels from the training data and 2096 from validation data, at a 9:1 ratio. Training began with an initial learning rate of 0.001, progressively reduced to 0.00001, using the Adam optimizer (https://www.geeksforgeeks.org/adam-optimizer/)¹⁸ to minimize the mean-squared-error loss function. We set the training for 100 epochs to ensure model convergence.

Phantom Experiments to Evaluate Noise and Spatial Resolution

A 20-cm-diameter American College of Radiology (ACR) CT accreditation phantom (Gammex; Sun Nuclear) was scanned to evaluate the noise power spectrum (NPS) and the contrast-dependent modulation transfer function (MTFc). All acquisitions and reconstructions were performed according to the clinical protocol settings used in this study. MTFc and NPS were calculated with data from Hr84-QIR3, Hr96-QIR3, and Hr96-CNN, using the online platform (https://www.ctpro.net) to illustrate the noise and resolution changes across different methods. In this study, NPS was computed by placing 10 square ROIs in the uniform section (module 3) of the ACR phantom. MTFc was calculated using bone cylindrical inserts (25-mm diameter, 4-cm depth) in module 1 of the phantom, on the basis of 40 consecutive axial slices, to assess in-plane spatial resolution under high-contrast conditions. A circular ROI was placed around the insert, and a circular-edge technique was used to measure the edge spread function by plotting the Hounsfield unit value of each pixel as a function of the distance from the center of the insert. The line spread function was then derived from the edge spread function. After zero padding, a fast Fourier transform was applied to the line spread function to compute the in-plane MTFc.

Objective Image-Quality Assessment

For image-quality evaluation, noise was measured in the axial images as the SD of CT numbers in a circular ROI drawn in a uniform soft-tissue area for each data set. The size and location of ROIs were matched among 3 image sets (Hr84-QIR3, Hr96-QIR3, and Hr96-CNN).

Reader Evaluation

Two fellowship-trained neuroradiologists (with >10 years’ experience each) assessed the overall image quality and delineation of 3 key anatomic structures—modiolus, stapes footplate, and incudomallear joint—for each of the 20 test cases. They assessed 3 image series per case (Hr84-QIR3, Hr96-QIR3, and Hr96-CNN), which were displayed side-by-side in a randomized and blinded manner. Images were ranked on a scale from 1 to 3, with 1 as the most preferred and 3 as the least preferred. Equal ranking was permitted.

Statistical Analyses

Statistical analyses were conducted using the statistical package scikit-posthocs in Python (https://scikit-posthocs.readthedocs.io/en/latest/). Pair-wise comparisons were performed with the Conover post hoc test, applying a Bonferroni correction, to evaluate differences between the 2 variables: the average rankings from 2 readers on overall image quality and diagnostic confidence for discerning 3 anatomic structures across Hr84-QIR3, Hr96-QIR3, and Hr96-CNN. A P value < .05 was considered statistically significant.

Denoising Performance Comparison

The residual encoder-decoder convolutional neural network (RED-CNN) and U-Net are 2 of the most widely used models for CT image denoising.¹⁹ Using our proposed data set preparation workflow, we conducted a comparative analysis of the denoising performance between RED-CNN and the U-Net model we developed. Hr96-QIR1 was used as the input reference, while Hr96-QIR3 served as the conventional denoised reconstruction. We compared the difference images generated by subtracting the reference noisy input image from the conventional method, RED-CNN, and U-Net denoised images. Image quality was objectively assessed using the structural similarity index (SSIM) and image noise measurements for each image set. This analysis was repeated at the patient level, with the mean (SD) of these metrics reported.