Source-linked AI summary
Can Deep Learning Outperform Modern Commercial CT Image Reconstruction Methods?
Hongming Shan, Atul Padole, Fatemeh Homayounieh, Uwe Kruger, Ruhani Doda Khera, Chayanin Nitiwarangkul, Mannudeep K. Kalra, Ge Wang
TL;DR
Commercial iterative reconstruction reduces LDCT noise but can raise concerns about image appearance and low-contrast detectability. This study introduces a progressive, radiologist-guided deep-learning denoising network and compares it with commercial IR across chest and abdominal scans from three vendors. The DL approach was comparable or favorable in noise suppression and structural fidelity and was much faster after training.
Problem
The study addresses whether deep neural networks can outperform modern commercial iterative reconstruction methods for LDCT while reducing dose-related image degradation.
Method
MAP-NN progressively denoises LDCT through repeated CLONE modules, learning intermediate targets and search gradients while allowing radiologists to select task-specific denoising depth.
Results
DL provided similar or better structural fidelity and noise suppression than commercial IR across three CT vendors and was much more computationally efficient after offline training.
Takeaways & Limitations
Deep-learning denoising can already compete with commercial IR and may potentially replace it, while vendor-agnostic processing may provide more consistent image appearance across institutions.
Takeaways & Limitations
MAP-NN was not optimized for a specific vendor or body region, and LDCT and NDCT slices were not perfectly registered.
Abstract
from arXiv · showhide
Commercial iterative reconstruction techniques on modern CT scanners target radiation dose reduction but there are lingering concerns over their impact on image appearance and low contrast detectability. Recently, machine learning, especially deep learning, has been actively investigated for CT. Here we design a novel neural network architecture for low-dose CT (LDCT) and compare it with commercial iterative reconstruction methods used for standard of care CT. While popular neural networks are trained for end-to-end mapping, driven by big data, our novel neural network is intended for end-to-process mapping so that intermediate image targets are obtained with the associated search gradients along which the final image targets are gradually reached. This learned dynamic process allows to include radiologists in the training loop to optimize the LDCT denoising workflow in a task-specific fashion with the denoising depth as a key parameter. Our progressive denoising network was trained with the Mayo LDCT Challenge Dataset, and tested on images of the chest and abdominal regions scanned on the CT scanners made by three leading CT vendors. The best deep learning based reconstructions are systematically compared to the best iterative reconstructions in a double-blinded reader study. It is found that our deep learning approach performs either comparably or favorably in terms of noise suppression and structural fidelity, and runs orders of magnitude faster than the commercial iterative CT reconstruction algorithms.
1 Introduction
The study asks whether deep learning can outperform commercial iterative reconstruction for LDCT while addressing dose-related noise and artifacts. It proposes progressive, radiologist-guided denoising rather than direct end-to-end mapping.
- Motivation: CT dose reduction can increase noise and artifacts, potentially compromising diagnostic performance.LDCT is used to reduce radiation-related risk, but untreated degradation can affect reconstructed-image quality.
- Related approaches: Existing LDCT noise-reduction methods include sinogram filtration, iterative reconstruction, and image post-processing.Sinogram methods can exploit known noise characteristics but may cause edge blurring or resolution loss.
- Study motivation: The study compares deep neural networks with modern commercial iterative reconstruction methods for LDCT across leading CT vendors.The authors emphasize comparing DL with vendor implementations rather than comparing image-quality metrics among vendors.
- Novel approach: Progressive denoising learns intermediate image targets and associated search gradients, allowing radiologists to guide task-specific denoising depth.This replaces conventional direct LDCT-to-NDCT mapping with an end-to-process paradigm.
- Network design: MAP-NN applies repeated CLONE modules to progressively denoise an LDCT FBP image.The formulation uses shared CLONE parameters across T stages, with g denoting a CLONE module and T the number of modules.
- Network design: CLONE extends an earlier LDCT denoising network with skip connections, output clipping, and modularization into a progressive model.The conveying links are intended to make the model more compact and flexible.
2 Material and Method
The study trained MAP-NN on paired Mayo low-dose and normal-dose abdominal CT data, then evaluated DL and commercial IR reconstructions in blinded reader assessments.
- Clinical data: The evaluation included 60 patients undergoing routine chest or abdominal CT on scanners from vendors A, B, and C.Half the datasets were abdominal and half chest examinations, with vendor representation proportional across the study.
- Reconstructions: LDCT sinograms were reconstructed using FBP and three selected clinical IR settings per vendor, while NDCT images used FBP.The LDCT FBP images served as inputs to the DL method.
- Reader study: Three radiologists independently reviewed randomized DL and IR images in a blinded fashion.Subjective image-quality assessments were performed for NDCT and LDCT cases independently.
- Training data: The training dataset contained normal-dose abdominal CT images from 10 patients and corresponding simulated quarter-dose images.Poisson noise was inserted into projection data to reach a noise level corresponding to 25% of normal dose.
- Training data: Training used 128K randomly selected 64 × 64 image patches from 5 patients, while testing used 64K patches from the remaining 5.The dataset came from the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge.
3 Results
Across vendors, body regions, and image-quality criteria, the best DL reconstructions were generally better than or comparable to the best IR reconstructions. DL also showed favorable lesion visibility and sample-image quality in several comparisons.
- Across vendors: DL was better than or comparable to IR overall, with vendors A and B favoring DL for abdominal imaging and comparability for chest imaging.For vendor C, readers found DL and IR comparable in both abdominal and chest regions.
- Image-quality criteria: DL outperformed IR in average structural fidelity for 12 of 18 classes, with equal performance in 3 additional classes.The classes span selected vendors and body regions.
- Image-quality criteria: DL was superior to IR in average noise suppression for 14 of 18 classes.In some cases, DL and IR had identical fidelity scores while DL achieved better noise suppression.
- Image-quality criteria: DL achieved significantly better noise suppression and structural fidelity than IR in one comparison, while other comparisons were better, comparable, or statistically comparable.Figure 5 reports mean scores and standard deviations for both criteria across vendor and region conditions.
- Sample images: Sample images showed better noise suppression and structural fidelity with DL for all vendors, and abdominal IR images from vendors A and B were judged unacceptable or limited.Those abdominal images were acceptable with MAP-NN for all three vendors.
- Lesion visibility: Two of 30 NDCT lesions were not seen on LDCT FBP, IR, or DL images, while the remaining 28 were seen equally well with IR and DL.A pseudo-lesion present on LDCT FBP and IR was absent from both DL and NDCT images.
4 Discussions
The study finds that MAP-NN achieves image quality comparable to or better than commercial iterative reconstruction across three vendors and two body regions, with faster post-training denoising.
- MAP-NN performs better than or comparably to clinically used iterative reconstruction methods across abdominal and chest CT from three vendors.
- After training, deep-learning denoising processes about 100 slices per second per mapping depth, whereas iterative reconstruction is time-consuming.
- MAP-NN systematically matches or exceeds commercial iterative reconstruction in structural fidelity and noise suppression.
- The vendor-agnostic approach could streamline multi-vendor radiomics by reducing vendor-specific differences in image appearance.
- The study was not optimized for a specific vendor or body region, and imperfect LDCT-NDCT registration may affect evaluation scores.
A.1 Conventional denoising model
The conventional denoising model treats LDCT reconstruction as learning an approximate inverse of a noise-corrupting process, typically mapping a specific low-dose image to its NDCT counterpart.
- The model represents an LDCT image as an NDCT image transformed by a noise-induced corrupting process N.
- Conventional denoising learns an approximate inverse function to estimate the NDCT image from the LDCT input.
- Existing implementations include fully connected convolutional, encoder-decoder, skip-connection, conveying-path, and three-dimensional network variants.
- Direct mapping from one dose level to normal dose may generalize poorly, requiring separate models for different radiation-dose settings.
- NDCT training targets still contain noise, unlike the nearly noise-free ground truth commonly used in computer-vision denoising.
A.2 Progressive denoising model
MAP-NN extends direct denoising into a modular progressive process that repeatedly applies a shared CLONE module, producing intermediate outputs for task-specific selection by radiologists.
- MAP-NN is a modular refinement of direct denoising that applies multiple CLONE stages progressively.
- T = 1 reproduces conventional denoising, while T > 1 increases depth and receptive-field size without adding new parameters.
- Each CLONE generates an intermediate denoised image, forming a dynamic sequence whose progression corresponds to increased effective radiation dose.
- Because each CLONE removes part of the noise, MAP-NN does not hypothetically require noise-free normal-dose labels.
- For a new dose level, repeated CLONE application produces candidate outputs that domain experts can select in the loop.
B MAP-NN architecture
MAP-NN uses repeated CLONE modules, while CLONE incorporates architectural modifications intended to stabilize progressive denoising training and preserve input information.
- MAP-NN contains multiple identical CLONE modules with shared parameters, and each CLONE can be based on an existing denoising network.
- The study’s CLONE extends the earlier CPCE encoder-decoder for progressive denoising.
- Training deep progressive denoising with CPCE alone is difficult because of exploding or vanishing gradients and information-copying demands.
- Output clipping to [0, 1] limits extreme outputs, while residual skip connections address vanishing gradients and let modules infer noise distributions.
B.3 Discriminator structure
The discriminator is used within a WGAN framework to distinguish generated denoised images from NDCT samples, while the generator is optimized with multiple loss components.
- Discriminator structure: The discriminator comprises six convolutional layers followed by two fully connected layers for distinguishing generated images from ground-truth samples.The convolutional layers use 64, 64, 128, 128, 256, and 256 filters, followed by fully connected layers of sizes 1024 and 1.
- Adversarial framework: The study uses a Wasserstein GAN with gradient penalty to address low-quality outputs, slow convergence, and mode collapse.The gradient penalty balances the Wasserstein distance and penalty term, with λ_g suggested to be 10.
- Adversarial objective: The adversarial loss encourages the generator to produce samples indistinguishable from NDCT images.The proposed MAP-NN serves as the generator G in the WGAN framework.
C.2 Mean-squared error
The training objective combines mean-squared error and edge information with adversarial loss to reduce noise while preserving texture and enhancing edges.
- Mean-squared error: Mean-squared error measures the difference between the denoised output and NDCT images, reducing noise in the input LDCT image.
- Edge information: The Sobel filter approximates image-intensity gradients, providing edge information for the denoising objective.The filter is computationally inexpensive but gives a relatively crude approximation for high-frequency variations.
- Final objective: The final generator objective combines adversarial loss, mean-squared error, and edge incoherence.The objective is designed to preserve texture information, reduce noise, and enhance edges.
D Training details
Training used Adam on image patches, while the study evaluated progressive denoising and reconstruction comparisons across vendors and anatomical regions.
- Optimization: The network was trained for 80 epochs within 24 hours on an NVIDIA 1080Ti GPU using TensorFlow.
- Progressive denoising: Figure 1 depicts MAP-NN modules and progressively denoised LDCT images, with mapping depth D controlling the denoising progression.Each module combines a skip connection, residual-learning encoder-decoder, summation, and output clipping.
- Reader study: Figure 2 compares the best DL and IR reconstructions across abdomen and chest scans from three vendors and three readers.The histograms classify 20 cases per vendor-region combination as DL better, equal, or worse than IR, with binomial-test significance at 5%.
- Image comparison: Figure 3 presents representative best DL, best IR, and NDCT FBP images for abdomen and chest scans from the three vendors.Colored arrows identify liver lesions, lung nodules, and centrilobular emphysema.
LDCT CNN NDCT f
The paper compares progressive and direct denoising architectures and evaluates deep-learning and iterative reconstructions across image-quality metrics, readers, vendors, and body regions.
- Model comparison: Progressive denoising uses multiple identical modules to generate intermediate results along the noise-reduction direction, whereas direct denoising learns one LDCT-to-NDCT mapping.With one module, the progressive model reduces to the direct model.
- Reconstruction comparison: 12 of 18 classes favored DL over IR for average structural fidelity, with equal performance in 3 additional classes.The comparison spans abdomen and chest scans from vendors A, B, and C.
- Reconstruction comparison: 14 of 18 classes favored DL over IR for average noise suppression across the evaluated body-region and vendor classes.The metrics include fidelity scores and noise suppression for LDCT, best DL, best IR, and NDCT.
- Reader assessment: Three-reader LDCT quality assessments showed Cohen’s kappa inter-reader agreement ranging from 0.42 to 0.70 for noise suppression and structural fidelity.The statistic summarizes agreement among the three readers.
- Training validation: Validation curves for abdomen and chest windows were trained for 80 epochs and converged on 64K Mayo clinical image patches.The figure presents abdomen and chest validation curves separately.