Image quality assessment of pediatric chest and abdomen CT by deep learning reconstruction

Low-tube-voltage whole-body CT angiography with extremely low iodine dose: a comparison between hybrid-iterative reconstruction and deep-learning image-reconstruction algorithms

AIM

To evaluate arterial enhancement, its depiction, and image quality in low-tube potential whole-body computed tomography (CT) angiography (CTA) with extremely low iodine dose and compare the results with those obtained by hybrid-iterative reconstruction (IR) and deep-learning image-reconstruction (DLIR) methods.

MATERIALS AND METHODS

This prospective study included 34 consecutive participants (27 men; mean age, 74.2 years) who underwent whole-body CTA at 80 kVp for evaluating aortic diseases between January and July 2020. Contrast material (240 mg iodine/ml) with simultaneous administration of its quarter volume of saline, which corresponded to 192 mg iodine/ml, was administered. CT raw data were reconstructed using adaptive statistical IR-Veo of 40% (hybrid-IR), DLIR with medium- (DLIR-M), and high-strength level (DLIR-H). A radiologist measured CT attenuation of the arteries and background noise, and the signal-to-noise ratio (SNR) was then calculated. Two reviewers qualitatively evaluated the arterial depictions and diagnostic acceptability on axial, multiplanar-reformatted (MPR), and volume-rendered (VR) images.

RESULTS

Mean contrast material volume and iodine weight administered were 64.1 ml and 15.4 g, respectively. The SNRs of the arteries were significantly higher in the following order of the DLIR-H, DLIR-M, and hybrid-IR (p<0.001). Depictions of six arteries on axial, three arteries on MPR, and four arteries on VR images were significantly superior in the DLIR-M or hybrid-IR than in the DLIR-H (p≤0.009 for each). Diagnostic acceptability was significantly better in the DLIR-M and DLIR-H than in the hybrid-IR (p<0.001-0.005).

CONCLUSION

DLIR-M showed well-balanced arterial depictions and image quality compared with the hybrid-IR and DLIR-H.

Low-dose liver CT: image quality and diagnostic accuracy of deep learning image reconstruction algorithm

OBJECTIVES

To perform a comprehensive within-subject image quality analysis of abdominal CT examinations reconstructed with DLIR and to evaluate diagnostic accuracy compared to the routinely applied adaptive statistical iterative reconstruction (ASiR-V) algorithm.

MATERIALS AND METHODS

Oncologic patients were prospectively enrolled and underwent contrast-enhanced CT. Images were reconstructed with DLIR with three intensity levels of reconstruction (high, medium, and low) and ASiR-V at strength levels from 10 to 100% with a 10% interval. Three radiologists characterized the lesions and two readers assessed diagnostic accuracy and calculated signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), figure of merit (FOM), and subjective image quality, the latter with a 5-point Likert scale.

RESULTS

Fifty patients (mean age: 70 ± 10 years, 23 men) were enrolled and 130 liver lesions (105 benign lesions, 25 metastases) were identified. DLIR_H achieved the highest SNR and CNR, comparable to ASiR-V 100% (p ≥ .051). DLIR_M returned the highest subjective image quality (score: 5; IQR: 4-5; p ≤ .001) and significant median increase (29%) in FOM (p < .001). Differences in detection were identified only for lesions ≤ 0.5 cm: 32/33 lesions were detected with DLIR_M and 26 lesions were detected with ASiR-V 50% (p = .031). Lesion accuracy of was 93.8% (95% CI: 88.1, 97.3; 122 of 130 lesions) for DLIR and 87.7% (95% CI: 80.8, 92.8; 114 of 130 lesions) for ASiR-V 50%.

CONCLUSIONS

DLIR yields superior image quality and provides higher diagnostic accuracy compared to ASiR-V in the assessment of hypovascular liver lesions, in particular for lesions ≤ 0.5 cm.

CLINICAL RELEVANCE STATEMENT

Deep learning image reconstruction algorithm demonstrates higher diagnostic accuracy compared to iterative reconstruction in the identification of hypovascular liver lesions, especially for lesions ≤ 0.5 cm.

KEY POINTS

• Iterative reconstruction algorithm impacts image texture, with negative effects on diagnostic capabilities. • Medium-strength deep learning image reconstruction algorithm outperforms iterative reconstruction in the diagnostic accuracy of ≤ 0.5 cm hypovascular liver lesions (93.9% vs 78.8%), also granting higher objective and subjective image quality. • Deep learning image reconstruction algorithm can be safely implemented in routine abdominal CT protocols in place of iterative reconstruction.

Deep learning-based algorithms for low-dose CT imaging: A review

The computed tomography (CT) technique is extensively employed as an imaging modality in clinical settings. The radiation dose of CT, however, is significantly high, thereby raising concerns regarding the potential radiation damage it may cause. The reduction of X-ray exposure dose in CT scanning may result in a significant decline in imaging quality, thereby elevating the risk of missed diagnosis and misdiagnosis. The reduction of CT radiation dose and acquisition of high-quality images to meet clinical diagnostic requirements have always been a critical research focus and challenge in the field of CT. Over the years, scholars have conducted extensive research on enhancing low-dose CT (LDCT) imaging algorithms, among which deep learning-based algorithms have demonstrated superior performance. In this review, we initially introduced the conventional algorithms for CT image reconstruction along with their respective advantages and disadvantages. Subsequently, we provided a detailed description of four aspects concerning the application of deep neural networks in LDCT imaging process: preprocessing in the projection domain, post-processing in the image domain, dual-domain processing imaging, and direct deep learning-based reconstruction (DLR). Furthermore, an analysis was conducted to evaluate the merits and demerits of each method. The commercial and clinical applications of the LDCT-DLR algorithm were also presented in an overview. Finally, we summarized the existing issues pertaining to LDCT-DLR and concluded the paper while outlining prospective trends for algorithmic advancement.

Prospective evaluation of deep learning image reconstruction for Lung-RADS and automatic nodule volumetry on ultralow-dose chest CT

PURPOSE

To prospectively evaluate whether Lung-RADS classification and volumetric nodule assessment were feasible with ultralow-dose (ULD) chest CT scans with deep learning image reconstruction (DLIR).

METHODS

The institutional review board approved this prospective study. This study included 40 patients (mean age, 66±12 years; 21 women). Participants sequentially underwent LDCT and ULDCT (CTDIvol, 0.96±0.15 mGy and 0.12±0.01 mGy) scans reconstructed with the adaptive statistical iterative reconstruction-V 50% (ASIR-V50) and DLIR. CT image quality was compared subjectively and objectively. The pulmonary nodules were assessed visually by two readers using the Lung-RADS 1.1 and automatically using a computerized assisted tool.

RESULTS

DLIR provided a significantly higher signal-to-noise ratio for LDCT and ULDCT images than ASIR-V50 (all P < .001). In general, DLIR showed superior subjective image quality for ULDCT images (P < .001) and comparable quality for LDCT images compared to ASIR-V50 (P = .01-1). The per-nodule sensitivities of observers for Lung-RADS category 3-4 nodules were 70.6-88.2% and 64.7-82.4% for DLIR-LDCT and DLIR-ULDCT images (P = 1) and categories were mostly concordant within observers. The per-nodule sensitivities of the computer-assisted detection for nodules ≥4 mm were 72.1% and 67.4% on DLIR-LDCT and ULDCT images (P = .50). The 95% limits of agreement for nodule volume differences between DLIR-LDCT and ULDCT images (-85.6 to 78.7 mm3) was similar to the within-scan nodule volume differences between DLIR- and ASIR-V50-LDCT images (-63.9 to 78.5 mm3), with volume differences smaller than 25% in 88.5% and 92.3% of nodules, respectively (P = .65).

CONCLUSION

DLIR enabled comparable Lung-RADS and volumetric nodule assessments on ULDCT images to LDCT images.

Towards safer imaging: A comparative study of deep learning-based denoising and iterative reconstruction in intraindividual low-dose CT scans using an in-vivo large animal model

PURPOSE

Computed tomography (CT) scans are a significant source of medically induced radiation exposure. Novel deep learning-based denoising (DLD) algorithms have been shown to enable diagnostic image quality at lower radiation doses than iterative reconstruction (IR) methods. However, most comparative studies employ low-dose simulations due to ethical constraints. We used real intraindividual animal scans to investigate the dose-reduction capabilities of a DLD algorithm in comparison to IR.

MATERIALS AND METHODS

Fourteen veterinarian-sedated alive pigs underwent 2 CT scans on the same 3rd generation dual-source scanner with two months between each scan. Four additional scans ensued each time, with mAs reduced to 50 %, 25 %, 10 %, and 5 %. All scans were reconstructed ADMIRE levels 2 (IR2) and a novel DLD algorithm, resulting in 280 datasets. Objective image quality (CT numbers stability, noise, and contrast-to-noise ratio) was measured via consistent regions of interest. Three radiologists independently rated all possible dataset combinations per time point for subjective image quality (-1 = inferior, 0 = equal, 1 = superior). The points were averaged for a semiquantitative score, and inter-rater agreement was measured using Spearman's correlation coefficient and adequately corrected mixed-effects modeling analyzed objective and subjective image quality.

RESULTS

Neither dose-reduction nor reconstruction method negatively impacted CT number stability (p > 0.999). In objective image quality assessment, the lowest radiation dose achievable by DLD when comparing noise (p = 0.544) and CNR (p = 0.115) to 100 % IR2 was 25 %. Overall, inter-rater agreement of the subjective image quality ratings was strong (r ≥ 0.69, mean 0.93 ± 0.05, 95 % CI 0.92-0.94; each p < 0.001), and subjective assessments corroborated that DLD at 25 % radiation dose was comparable to 100 % IR2 in image quality, sharpness, and contrast (p ≥ 0.281).

CONCLUSIONS

The DLD algorithm can achieve image quality comparable to the standard IR method but with a significant dose reduction of up to 75%. This suggests a promising avenue for lowering patient radiation exposure without sacrificing diagnostic quality.

CT image denoising methods for image quality improvement and radiation dose reduction

With the ever-increasing use of computed tomography (CT), concerns about its radiation dose have become a significant public issue. To address the need for radiation dose reduction, CT denoising methods have been widely investigated and applied in low-dose CT images. Numerous noise reduction algorithms have emerged, such as iterative reconstruction and most recently, deep learning (DL)-based approaches. Given the rapid advancements in Artificial Intelligence techniques, we recognize the need for a comprehensive review that emphasizes the most recently developed methods. Hence, we have performed a thorough analysis of existing literature to provide such a review. Beyond directly comparing the performance, we focus on pivotal aspects, including model training, validation, testing, generalizability, vulnerability, and evaluation methods. This review is expected to raise awareness of the various facets involved in CT image denoising and the specific challenges in developing DL-based models.

Generative Adversarial Network (Generative Artificial Intelligence) in Pediatric Radiology: A Systematic Review

Generative artificial intelligence, especially with regard to the generative adversarial network (GAN), is an important research area in radiology as evidenced by a number of literature reviews on the role of GAN in radiology published in the last few years. However, no review article about GAN in pediatric radiology has been published yet. The purpose of this paper is to systematically review applications of GAN in pediatric radiology, their performances, and methods for their performance evaluation. Electronic databases were used for a literature search on 6 April 2023. Thirty-seven papers met the selection criteria and were included. This review reveals that the GAN can be applied to magnetic resonance imaging, X-ray, computed tomography, ultrasound and positron emission tomography for image translation, segmentation, reconstruction, quality assessment, synthesis and data augmentation, and disease diagnosis. About 80% of the included studies compared their GAN model performances with those of other approaches and indicated that their GAN models outperformed the others by 0.1-158.6%. However, these study findings should be used with caution because of a number of methodological weaknesses. For future GAN studies, more robust methods will be essential for addressing these issues. Otherwise, this would affect the clinical adoption of the GAN-based applications in pediatric radiology and the potential advantages of GAN could not be realized widely.

Deep-learning CT reconstruction in clinical scans of the abdomen: a systematic review and meta-analysis

OBJECTIVE

To perform a systematic literature review and meta-analysis of the two most common commercially available deep-learning algorithms for CT.

METHODS

We used PubMed, Scopus, Embase, and Web of Science to conduct systematic searches for studies assessing the most common commercially available deep-learning CT reconstruction algorithms: True Fidelity (TF) and Advanced intelligent Clear-IQ Engine (AiCE) in the abdomen of human participants since only these two algorithms currently have adequate published data for robust systematic analysis.

RESULTS

Forty-four articles fulfilled inclusion criteria. 32 studies evaluated TF and 12 studies assessed AiCE. DLR algorithms produced images with significantly less noise (22-57.3% less than IR) but preserved a desirable noise texture with increased contrast-to-noise ratios and improved lesion detectability on conventional CT. These improvements with DLR were similarly noted in dual-energy CT which was only assessed for a single vendor. Reported radiation reduction potential was 35.1-78.5%. Nine studies assessed observer performance with the two dedicated liver lesion studies being performed on the same vendor reconstruction (TF). These two studies indicate preserved low contrast liver lesion detection (> 5 mm) at CTDIvol 6.8 mGy (BMI 23.5 kg/m2) to 12.2 mGy (BMI 29 kg/m2). If smaller lesion detection and improved lesion characterization is needed, a CTDIvol of 13.6-34.9 mGy is needed in a normal weight to obese population. Mild signal loss and blurring have been reported at high DLR reconstruction strengths.

CONCLUSION

Deep learning reconstructions significantly improve image quality in CT of the abdomen. Assessment of other dose levels and clinical indications is needed. Careful choice of radiation dose levels is necessary, particularly for small liver lesion assessment.

Computed Tomography 2.0: New Detector Technology, AI, and Other Developments

ABSTRACT

Computed tomography (CT) dramatically improved the capabilities of diagnostic and interventional radiology. Starting in the early 1970s, this imaging modality is still evolving, although tremendous improvements in scan speed, volume coverage, spatial and soft tissue resolution, as well as dose reduction have been achieved. Tube current modulation, automated exposure control, anatomy-based tube voltage (kV) selection, advanced x-ray beam filtration, and iterative image reconstruction techniques improved image quality and decreased radiation exposure. Cardiac imaging triggered the demand for high temporal resolution, volume acquisition, and high pitch modes with electrocardiogram synchronization. Plaque imaging in cardiac CT as well as lung and bone imaging demand for high spatial resolution. Today, we see a transition of photon-counting detectors from experimental and research prototype setups into commercially available systems integrated in patient care. Moreover, with respect to CT technology and CT image formation, artificial intelligence is increasingly used in patient positioning, protocol adjustment, and image reconstruction, but also in image preprocessing or postprocessing. The aim of this article is to give an overview of the technical specifications of up-to-date available whole-body and dedicated CT systems, as well as hardware and software innovations for CT systems in the near future.

Coronary Computed Tomography Angiography with Deep Learning Image Reconstruction: A Preliminary Study to Evaluate Radiation Exposure Reduction

Coronary computed tomography angiography (CCTA) is a medical imaging technique that produces detailed images of the coronary arteries. Our work focuses on the optimization of the prospectively ECG-triggered scan technique, which delivers the radiation efficiently only during a fraction of the R-R interval, matching the aim of reducing radiation dose in this increasingly used radiological examination. In this work, we analyzed how the median DLP (Dose-Length Product) values for CCTA of our Center decreased significantly in recent times mainly due to a notable change in the technology used. We passed from a median DLP value of 1158 mGy·cm to 221 mGy·cm for the whole exam and from a value of 1140 mGy·cm to 204 mGy·cm if considering CCTA scanning only. The result was obtained through the association of important factors during the dose imaging optimization: technological improvement, acquisition technique, and image reconstruction algorithm intervention. The combination of these three factors allows us to perform a faster and more accurate prospective CCTA with a lower radiation dose. Our future aim is to tune the image quality through a detectability-based study, combining algorithm strength with automatic dose settings.

Improving detection performance of hepatocellular carcinoma and interobserver agreement for liver imaging reporting and data system on CT using deep learning reconstruction

PURPOSE

This study aimed to compare the hepatocellular carcinoma (HCC) detection performance, interobserver agreement for Liver Imaging Reporting and Data System (LI-RADS) categories, and image quality between deep learning reconstruction (DLR) and conventional hybrid iterative reconstruction (Hybrid IR) in CT.

METHODS

This retrospective study included patients who underwent abdominal dynamic contrast-enhanced CT between October 2021 and March 2022. Arterial, portal, and delayed phase images were reconstructed using DLR and Hybrid IR. Two blinded readers independently read the image sets with detecting HCCs, scoring LI-RADS, and evaluating image quality.

RESULTS

A total of 26 patients with HCC (mean age, 73 years ± 12.3) and 23 patients without HCC (mean age, 66 years ± 14.7) were included. The figures of merit (FOM) for the jackknife alternative free-response receiver operating characteristic analysis in detecting HCC averaged for the readers were 0.925 (reader 1, 0.937; reader 2, 0.913) in DLR and 0.878 (reader 1, 0.904; reader 2, 0.851) in Hybrid IR, and the FOM in DLR were significantly higher than that in Hybrid IR (p = 0.038). The interobserver agreement (Cohen's weighted kappa statistics) for LI-RADS categories was moderate for DLR (0.595; 95% CI, 0.585-0.605) and significantly superior to Hybrid IR (0.568; 95% CI, 0.553-0.582). According to both readers, DLR was significantly superior to Hybrid IR in terms of image quality (p ≤ 0.021).

CONCLUSION

DLR improved HCC detection, interobserver agreement for LI-RADS categories, and image quality in evaluations of HCC compared to Hybrid IR in abdominal dynamic contrast-enhanced CT.

CT image quality evaluation in the age of deep learning: trade-off between functionality and fidelity

OBJECTIVE

To quantitatively compare DLIR and ASiR-V with realistic anatomical images.

METHODS

CT scans of an anthropomorphic phantom were acquired using three routine protocols (brain, chest, and abdomen) at four dose levels, with images reconstructed at five levels of ASiR-V and three levels of DLIR. Noise power spectrum (NPS) was estimated using a difference image by subtracting two matching images from repeated scans. Using the max-dose FBP reconstruction as the ground truth, the structure similarity index (SSIM) and gradient magnitude (GM) of difference images were evaluated. Image noise magnitude (σ), frequency location of the NPS peak (f_peak), mean SSIM (MSSIM), and mean GM (MGM) were used as quantitative metrics to compare image quality, for each anatomical region, protocol, algorithm, dose level, and slice thickness.

RESULTS

Image noise had a strong (R² > 0.99) power law relationship with dose for all algorithms. For the abdomen and chest, f_peak shifted from 0.3 (FBP) down to 0.15 mm^-1 (ASiR-V 100%) with increasing ASiR-V strength but remained 0.3 mm^-1 for all DLIR levels. f_peak shifted down for the brain protocol with increasing DLIR levels. Three levels of DLIR produced similar image noise levels as ASiR-V 40%, 80%, and 100%, respectively. DLIR had lower MSSIM but higher MGM than ASiR-V while matching imaging noise.

CONCLUSION

Compared to ASiR-V, DLIR presents trade-offs between functionality and fidelity: it has a noise texture closer to FBP and more edge enhancement, but reduced structure similarity. These trade-offs and unique protocol-dependent behaviors of DLIR should be considered during clinical implementation and deployment.

KEY POINTS

• DLIR reconstructed images demonstrate closer noise texture and lower structure similarity to FBP while producing equivalent noise levels comparable to ASiR-V. • DLIR has additional edge enhancement as compared to ASiR-V. • DLIR has unique protocol-dependent behaviors that should be considered for clinical implementation.

Deep Learning Image Reconstruction for CT: Technical Principles and Clinical Prospects

Filtered back projection (FBP) has been the standard CT image reconstruction method for 4 decades. A simple, fast, and reliable technique, FBP has delivered high-quality images in several clinical applications. However, with faster and more advanced CT scanners, FBP has become increasingly obsolete. Higher image noise and more artifacts are especially noticeable in lower-dose CT imaging using FBP. This performance gap was partly addressed by model-based iterative reconstruction (MBIR). Yet, its "plastic" image appearance and long reconstruction times have limited widespread application. Hybrid iterative reconstruction partially addressed these limitations by blending FBP with MBIR and is currently the state-of-the-art reconstruction technique. In the past 5 years, deep learning reconstruction (DLR) techniques have become increasingly popular. DLR uses artificial intelligence to reconstruct high-quality images from lower-dose CT faster than MBIR. However, the performance of DLR algorithms relies on the quality of data used for model training. Higher-quality training data will become available with photon-counting CT scanners. At the same time, spectral data would greatly benefit from the computational abilities of DLR. This review presents an overview of the principles, technical approaches, and clinical applications of DLR, including metal artifact reduction algorithms. In addition, emerging applications and prospects are discussed.

Adaptation of Deep Learning Image Reconstruction for Pediatric Head CT: A Focus on the Image Quality

Purpose

To assess the effect of deep learning image reconstruction (DLIR) for head CT in pediatric patients.

Materials and Methods

We collected 126 pediatric head CT images, which were reconstructed using filtered back projection, iterative reconstruction using adaptive statistical iterative reconstruction (ASiR)-V, and all three levels of DLIR (TrueFidelity; GE Healthcare). Each image set group was divided into four subgroups according to the patients' ages. Clinical and dose-related data were reviewed. Quantitative parameters, including the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), and qualitative parameters, including noise, gray matter-white matter (GM-WM) differentiation, sharpness, artifact, acceptability, and unfamiliar texture change were evaluated and compared.

Results

The SNR and CNR of each level in each age group increased among strength levels of DLIR. High-level DLIR showed a significantly improved SNR and CNR (p < 0.05). Sequential reduction of noise, improvement of GM-WM differentiation, and improvement of sharpness was noted among strength levels of DLIR. Those of high-level DLIR showed a similar value as that with ASiR-V. Artifact and acceptability did not show a significant difference among the adapted levels of DLIR.

Conclusion

Adaptation of high-level DLIR for the pediatric head CT can significantly reduce image noise. Modification is needed while processing artifacts.

Deep learning versus iterative reconstruction on image quality and dose reduction in abdominal CT: a live animal study

Objective. While simulated low-dose CT images and phantom studies cannot fully approximate subjective and objective effects of deep learning (DL) denoising on image quality, live animal models may afford this assessment. This study is to investigate the potential of DL in CT dose reduction on image quality compared to iterative reconstruction (IR). Approach. The upper abdomen of a live 4 year old sheep was scanned on a CT scanner at different exposure levels. Images were reconstructed using FBP and ADMIRE with 5 strengths. A modularized DL network with 5 modules was used for image reconstruction via progressive denoising. Radiomic features were extracted from a region over the liver. Concordance correlation coefficient (CCC) was applied to quantify agreement between any two sets of radiomic features. Coefficient of variation was calculated to measure variation in a radiomic feature series. Structural similarity index (SSIM) was used to measure the similarity between any two images. Diagnostic quality, low-contrast detectability, and image texture were qualitatively evaluated by two radiologists. Pearson correlation coefficient was computed across all dose-reconstruction/denoising combinations. Results. A total of 66 image sets, with 405 radiomic features extracted from each, are analyzed. IR and DL can improve diagnostic quality and low-contrast detectability and similarly modulate image texture features. In terms of SSIM, DL has higher potential in preserving image structure. There is strong correlation between SSIM and radiologists’ evaluations for diagnostic quality (0.559) and low-contrast detectability (0.635) but moderate correlation for texture (0.313). There is moderate correlation between CCC of radiomic features and radiologists’ evaluation for diagnostic quality (0.397), low-contrast detectability (0.417), and texture (0.326), implying that improvement of image features may not relate to improvement of diagnostic quality. Conclusion. DL shows potential to further reduce radiation dose while preserving structural similarity, while IR is favored by radiologists and more predictably alters radiomic features.

Artificial Intelligence for Radiation Dose Optimization in Pediatric Radiology: A Systematic Review

Radiation dose optimization is particularly important in pediatric radiology, as children are more susceptible to potential harmful effects of ionizing radiation. However, only one narrative review about artificial intelligence (AI) for dose optimization in pediatric computed tomography (CT) has been published yet. The purpose of this systematic review is to answer the question “What are the AI techniques and architectures introduced in pediatric radiology for dose optimization, their specific application areas, and performances?” Literature search with use of electronic databases was conducted on 3 June 2022. Sixteen articles that met selection criteria were included. The included studies showed deep convolutional neural network (CNN) was the most common AI technique and architecture used for dose optimization in pediatric radiology. All but three included studies evaluated AI performance in dose optimization of abdomen, chest, head, neck, and pelvis CT; CT angiography; and dual-energy CT through deep learning image reconstruction. Most studies demonstrated that AI could reduce radiation dose by 36–70% without losing diagnostic information. Despite the dominance of commercially available AI models based on deep CNN with promising outcomes, homegrown models could provide comparable performances. Future exploration of AI value for dose optimization in pediatric radiology is necessary due to small sample sizes and narrow scopes (only three modalities, CT, positron emission tomography/magnetic resonance imaging and mobile radiography, and not all examination types covered) of existing studies.

WBC image classification and generative models based on convolutional neural network

Background

Computer-aided methods for analyzing white blood cells (WBC) are popular due to the complexity of the manual alternatives. Recent works have shown highly accurate segmentation and detection of white blood cells from microscopic blood images. However, the classification of the observed cells is still a challenge, in part due to the distribution of the five types that affect the condition of the immune system.

Methods

(i) This work proposes W-Net, a CNN-based method for WBC classification. We evaluate W-Net on a real-world large-scale dataset that includes 6562 real images of the five WBC types. (ii) For further benefits, we generate synthetic WBC images using Generative Adversarial Network to be used for education and research purposes through sharing.

Results

(i) W-Net achieves an average accuracy of 97%. In comparison to state-of-the-art methods in the field of WBC classification, we show that W-Net outperforms other CNN- and RNN-based model architectures. Moreover, we show the benefits of using pre-trained W-Net in a transfer learning context when fine-tuned to specific task or accommodating another dataset. (ii) The synthetic WBC images are confirmed by experiments and a domain expert to have a high degree of similarity to the original images. The pre-trained W-Net and the generated WBC dataset are available for the community to facilitate reproducibility and follow up research work.

Conclusion

This work proposed W-Net, a CNN-based architecture with a small number of layers, to accurately classify the five WBC types. We evaluated W-Net on a real-world large-scale dataset and addressed several challenges such as the transfer learning property and the class imbalance. W-Net achieved an average classification accuracy of 97%. We synthesized a dataset of new WBC image samples using DCGAN, which we released to the public for education and research purposes.