CheSS: Chest X-Ray Pre-trained Model via Self-supervised Contrastive Learning.

Training deep learning models on medical images heavily depends on experts' expensive and laborious manual labels. In addition, these images, labels, and even models themselves are not widely publicly accessible and suffer from various kinds of bias and imbalances. In this paper, chest X-ray pre-trained model via self-supervised contrastive learning (CheSS) was proposed to learn models with various representations in chest radiographs (CXRs). Our contribution is a publicly accessible pretrained model trained with a 4.8-M CXR dataset using self-supervised learning with a contrastive learning and its validation with various kinds of downstream tasks including classification on the 6-class diseases in internal dataset, diseases classification in CheXpert, bone suppression, and nodule generation. When compared to a scratch model, on the 6-class classification test dataset, we achieved 28.5% increase in accuracy. On the CheXpert dataset, we achieved 1.3% increase in mean area under the receiver operating characteristic curve on the full dataset and 11.4% increase only using 1% data in stress test manner. On bone suppression with perceptual loss, we achieved improvement in peak signal to noise ratio from 34.99 to 37.77, structural similarity index measure from 0.976 to 0.977, and root-square-mean error from 4.410 to 3.301 when compared to ImageNet pretrained model. Finally, on nodule generation, we achieved improvement in Fréchet inception distance from 24.06 to 17.07. Our study showed the decent transferability of CheSS weights. CheSS weights can help researchers overcome data imbalance, data shortage, and inaccessibility of medical image datasets. CheSS weight is available at https://github.com/mi2rl/CheSS .

Value of bone suppression software in chest radiographs for improving image quality and reducing radiation dose.

OBJECTIVES: To compare image quality and radiation dose between dual-energy subtraction (DES)-based bone suppression images (D-BSIs) and software-based bone suppression images (S-BSIs). METHODS: Chest radiographs (CXRs) of forty adult patients were obtained with the two X-ray devices, one with DES and one with bone suppression software. Three image quality metrics (relative mean absolute error (RMAE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM)) between original CXR and BSI for each of D-BSI and S-SBI groups were calculated for each bone and soft tissue areas. Two readers rated the visual image quality for original CXR and BSI for each of D-BSI and S-SBI groups. The dose area product (DAP) values were recorded. Paired t test was used to compare the image quality and DAP values between D-BSI and S-BSI groups. RESULTS: In bone areas, S-BSIs had better SSIM values than D-BSI (94.57 vs. 87.77) but worse RMAE and PSNR values (0.50 vs. 0.20; 20.93 vs. 34.37) (all p < 0.001). In soft tissue areas, S-BSIs had better SSIM values than D-BSI (97.56 vs. 91.42) but similar RMAE and PSNR values (0.29 vs. 0.27; 31.35 vs. 29.87) (all p < 0.001). Both readers gave S-BSIs significantly higher image quality scores than D-BSI (p < 0.001). The mean DAP in software-related images (0.98 dGy·cm2) was significantly lower than that in the DES-related images (1.48 dGy·cm2) (p < 0.001). CONCLUSION: Bone suppression software significantly improved the image quality of bone suppression images with a relatively lower radiation dose, compared with dual-energy subtraction technique. KEY POINTS: • Bone suppression software preserves structure similarity of soft tissues better than dual-energy subtraction technique in bone suppression images. • Bone suppression software achieves superior image quality for lung lesions than dual-energy subtraction technique in bone suppression images. • Bone suppression software can decrease the radiation dose over the hardware-based image processing technique.

Added Value of Bone Suppression Image in the Detection of Subtle Lung Lesions on Chest Radiographs with Regard to Reader's Expertise.

BACKGROUND: Chest radiographs (CXR) are the most commonly used imaging techniques by various clinicians and radiologists. However, detecting lung lesions on CXR depends largely on the reader's experience level, so there have been several trials to overcome this problem using post-processing of CXR. We investigated the added value of bone suppression image (BSI) in detecting various subtle lung lesions on CXR with regard to reader's expertise. METHODS: We applied a software program to generate BSI in 1,600 patients in the emergency department. Of them, 80 patients with subtle lung lesions and 80 patients with negative finding on CXR were retrospectively selected based on the subtlety scores on CXR and CT findings. Ten readers independently rated their confidence in deciding the presence or absence of a lung lesion at each of 960 lung regions on the two separated imaging sessions: CXR alone vs. CXR with BSI. RESULTS: The additional use of BSI for all readers significantly increased the mean area under the curve (AUC) in detecting subtle lung lesions (0.663 vs. 0.706; P < 0.001). The less experienced readers were, the more AUC differences increased: 0.067 (P < 0.001) for junior radiology residents; 0.064 (P < 0.001) for non-radiology clinicians; 0.044 (P < 0.001) for senior radiology residents; and 0.019 (P = 0.041) for chest radiologists. The additional use of BSI significantly increased the mean confidence regarding the presence or absence of lung lesions for 213 positive lung regions (2.083 vs. 2.357; P < 0.001) and for 747 negative regions (1.217 vs. 1.195; P = 0.008). CONCLUSION: The use of BSI increases diagnostic performance and confidence, regardless of reader's expertise, reduces the impact of reader's expertise and can be helpful for less experienced clinicians and residents in the detection of subtle lung lesions.

Intraductal Papillary Neoplasm of the Bile Duct: Clinical, Imaging, and Pathologic Features.

OBJECTIVE: We outline the concept of intraductal papillary neoplasm of the bile duct (IPNB), discuss the morphologic features of IPNB and the differential diagnoses, and describe the radiologic approaches used in multidisciplinary management. CONCLUSION: The concept of IPNB has been evolving. Because the imaging features of IPNB can be variable, different mimickers according to IPNB subtype can be considered. A multimodality approach is essential to obtain an optimal diagnosis and establish treatment plans.

Diffusion-weighted imaging with reverse phase-encoding polarity: the added value to the conventional diffusion-weighted imaging in differentiating acute infarctions from hyperintense brainstem artifacts.

OBJECTIVES: The aim of our study was to assess the value of diffusion-weighted imaging with reverse phase-encoding polarity (R-DWI) in addition to conventional DWI using forward phase-encoding polarity (F-DWI) in differentiating acute brainstem infarctions from hyperintense artefacts. METHODS: Seventy-six patients with 38 hyperintense brainstem artefacts and 38 acute brainstem infarctions that had undergone F-DWI and R-DWI were retrospectively selected based on the clinicoradiological diagnosis. Four radiologists independently rated their confidence in diagnosing acute infarctions and ruling out brainstem artefacts in a blind manner, and then compared the diagnostic performance and confidence between F-DWI alone and F-DWI with R-DWI. RESULTS: The areas under the curve determined for F-DWI with R-DWI in diagnosing infarctions were significantly higher than F-DWI alone for all readers (resident 1, 0.908 vs 0.776; resident 2, 0.908 vs 0.789; neuroradiologist, 0.961 vs 0.868; emergency radiologist, 0.934 vs 0.855, all p < 0.05). All readers were more confident using F-DWI with R-DWI than F-DWI alone (all p < 0.05) for diagnosing acute brainstem infarction, and three readers (readers except the neuroradiologist) were more confident using F-DWI with R-DWI for ruling out brainstem artefacts (p ≤ 0.001). CONCLUSION: The addition of R-DWI to F-DWI is a valuable method for differentiating acute brainstem infarctions from hyperintense artefacts. KEY POINTS: • Hyperintense brainstem artefacts can be confused with acute infarctions on DWI. • Additional R-DWI to F-DWI reduces inter-reader variability in diagnosing brainstem infarctions. • Additional R-DWI improves performance and confidence for discriminating infarctions from artefacts.

Apparent diffusion coefficient parametric response mapping MRI for follow-up of glioblastoma.

OBJECTIVES: To determine the diagnostic superiority of parametric response mapping of apparent diffusion coefficient (ADCPR) for predicting glioblastoma treatment response, compared to single time point measurement. METHODS: Fifty post-treatment glioblastoma patients were enrolled. ADCPR was calculated from serial apparent diffusion coefficient (ADC) maps acquired before and at the time of first detection of an enlarged contrast-enhancing lesion on voxel-by-voxel basis. The percentage-decrease in ADCPR and tenth percentile histogram cutoff value of ADC (ADC10) were compared at subsequent 3-month and 1-year follow-ups. RESULTS: The percentage-decrease in ADCPR was significantly higher in the progression group (mean = 33.2-38.3 %) than in the stable-response group (mean = 9.7 %) at 3 months follow-up (corrected p < 0.001 for both readers). ADCPR significantly improved area under the receiver operating characteristic curve from 0.67 to 0.88 (corrected p = 0.037) and from 0.70 to 0.92 (corrected p = 0.020) for both readers, respectively, compared to ADC10 at 3-month follow-up, but did not significantly improve at 1-year follow-up. The inter-reader agreement was higher for ADCPR than ADC10 (intraclass correlation coefficient, 0.93 versus 0.86). CONCLUSION: Voxel-based ADCPR appears to be a superior imaging biomarker than ADC, particularly for predicting early tumour progression in patients with glioblastoma. KEY POINTS: • Treatment response pattern of glioblastoma was evaluated using voxel-based ADCPR and ADC10. • Voxel-based ADCPR was more accurate in predicting treatment response pattern than ADC10. • Inter-reader agreement was higher in ADCPR calculation than in ADC10 calculation. • Voxel-based ADCPR can be a predictor of early treatment response pattern for glioblastoma.

Thread sign in biliary intraductal papillary mucinous neoplasm: a novel specific finding for MRI.

OBJECTIVES: To evaluate thread sign of biliary intraductal papillary mucinous neoplasm (B-IPMN) on magnetic resonance imaging (MRI). METHODS: Thread sign was defined as intraductal linear or curvilinear hypointense striations. Two radiologists independently evaluated the presence and location of thread sign on MR cholangiography (thin-slice, thick-slab and 3D MRC) and axial MR images (T2 TSE, T2 HASTE and DWI) in patients with B-IPMN (n = 38) and in matched control groups with benign (n = 36) or malignant (n = 35) biliary diseases. Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of thread sign for diagnosing B-IPMN were evaluated. RESULTS: Thread sign was observed only in patients with B-IPMN on MRC (44.7-52.6 % [17/38-20/38], P < 0.001) and axial MR images (31.6 % [12/38], P < 0.001), except in one patient with recurrent pyogenic cholangitis on MRC (2.8 %, 1/36). The sensitivity, specificity, accuracy, PPV and NPV of thread sign for diagnosing B-IPMN on MRC were 0.53, 0.99, 0.83, 0.95 and 0.80, respectively (reader 1) and 0.45, 1.0, 0.81, 1.0 and 0.77, respectively (reader 2). Thread sign was detected mainly at the extrahepatic bile duct (52.6 %, 20/38). CONCLUSION: B-IPMN can manifest thread sign, a novel specific MR finding, mainly at the extrahepatic bile duct on MRI, especially on MRC. KEY POINTS: • Some B-IPMNs manifest thread sign within the bile ducts on MRI. • Thread sign is a highly specific finding for B-IPMN on MRI. • MRC is superior to axial T2WI and DWI for detecting thread sign.

Iohexol versus diatrizoate for fecal/fluid tagging during CT colonography performed with cathartic preparation: comparison of examination quality.

OBJECTIVE: We aimed to compare iohexol vs. diatrizoate as fecal/fluid tagging agents for computed tomography colonography (CTC) regarding examination quality. METHODS: Forty prospective patients (M:F = 23:17; 63 ± 11.6 years) received CTC using 50 mL (350 mgI/mL) oral iohexol for tagging. Forty other indication-matched, age-matched, and sex-matched patients who underwent CTC using 100 mL diatrizoate for tagging and otherwise the same technique, were retrospectively identified. Two groups were compared regarding overall examination quality, per-patient and per-segment scores of colonic bubbles (0 [no bubbles] to 5 [the largest amount]), and the volume, attenuation, and homogeneity (untagged, layered, and homogeneous) of the residual colonic fluid. RESULTS: The iohexol group demonstrated a greater amount of colonic bubbles than the diatrizoate group: mean per-patient scores ± SD of 1.2 ± 0.8 vs. 0.7 ± 0.6, respectively (p = 0.003); and rates of segments showing ≥ grade 3 bubbles of 12.9 % (85/659) vs. 1.6 % (11/695), respectively (p = 0.001). Residual colonic fluid amount standardized to the colonic volume did not significantly differ: 7.2 % ± 4.2 vs. 7.8 % ± 3.7, respectively (p = 0.544). Tagged fluid attenuation was mostly comparable between groups and the fluid was homogeneously tagged in 98.7 % (224/227) vs. 99.5 % (218/219) segments, respectively (p = 0.344). Iohexol caused more colonic bubbles when used during cathartic CTC. Otherwise, examination quality was similarly adequate with both iohexol and diatrizoate. KEY POINTS: • When used for tagging, iohexol caused significantly more colonic bubbles than diatrizoate. • The residual colonic fluid amount did not significantly differ between iohexol and diatrizoate. • The quality of fluid tagging was similarly adequate in both iohexol and diatrizoate.

Simethicone to prevent colonic bubbles during CT colonography performed with polyethylene glycol lavage and iohexol tagging: a randomized clinical trial.

OBJECTIVE: The purpose of this study was to determine whether the occurrence of numerous colonic bubbles during CT colonography (CTC) performed with polyethylene glycol cleansing and oral iohexol fecal/fluid tagging could be prevented by use of simethicone. SUBJECTS AND METHODS: Adults with suspected colonic neoplasia who had been randomly assigned to control and simethicone intervention groups underwent CTC after cleansing with 4 L of polyethylene glycol, tagging with 50 mL of 350 mg I/mL oral iohexol, and without (control) or with (intervention) oral administration of 200 mg of simethicone. Colonic segments in the control and intervention groups were evaluated for amount of colonic bubbles during CTC. A 6-point grading system was used in which 0 indicated no bubbles and 5 indicated that more than three fourths of the air-distended mucosa was covered with bubbles. The primary endpoint was a per-patient colonic bubble grade, derived as an average of the segmental grades. RESULTS: Eighty adults with suspected colonic neoplasia were randomly assigned to the control (40 patients) and simethicone intervention (40 patients) groups. A total of 659 colonic segments in the control group and 689 segments in the intervention group were evaluated for amount of colonic bubbles during CTC. The per-patient colonic bubble score was significantly lower in the simethicone intervention group than in the control group. The mean score was 0.0±0.1 (SD) versus 1.2±0.8 (p<0.001; 95% CI for the mean difference, -1.4 to -1.0). In the intervention group, 673 (97.7%) segments were grade 0, and 16 (2.3%) were grade 1. In contrast, in the control group, 226 (34.3%) segments were grade 0; 173 (26.3%), grade 1; 175 (26.6%), grade 2; 45 (6.8%), grade 3; 23 (3.5%), grade 4; and 17 (2.6%), grade 5. CONCLUSION: The colonic bubbles associated with fecal/fluid tagging with iohexol can be successfully prevented by adding simethicone to the colonic preparation.

Differentiation of small (≤2 cm) hepatocellular carcinomas from small benign nodules in cirrhotic liver on gadoxetic acid-enhanced and diffusion-weighted magnetic resonance images.

OBJECTIVE: To identify imaging characteristics that differentiate small (≤2 cm) HCCs from small (≤2 cm) benign nodules in cirrhotic liver on gadoxetic acid-enhanced and diffusion-weighted (DW) magnetic resonance (MR) images. MATERIALS AND METHODS: On gadoxetic acid-enhanced and DW MR images, we analysed signal intensity of 222 small HCCs and 61 benign nodules (diameter, 0.5-2 cm) at each sequence and rim enhancement during portal or equilibrium phases. Univariate and multivariate logistic regression analyses identified predictors of HCC. Combinations of significant MR findings in multivariate analysis were compared with American Association for the Study of Liver Disease (AASLD) practice guidelines. RESULTS: In multivariate analysis, arterial enhancement (adjusted odds ratio [aOR], 8.6), T2 hyperintensity (aOR, 5.8), and hyperintensity on DW images (aOR, 3.8) were significant for differentiating small HCCs from benign nodules (p ≤ 0.004). When two or all three findings were applied as diagnostic criteria for differentiating small HCCs from benign nodules, sensitivity and accuracy were significantly higher compared with AASLD practice guidelines (91% vs. 78% and 89% vs. 81%, respectively; each p < 0.0001). CONCLUSION: On gadoxetic acid-enhanced MR imaging, arterial enhancement and hyperintensity on T2-weighted and DW MR images are helpful for differentiating small HCCs from benign nodules in liver cirrhosis.

Lateral orbital wall fracture as a mechanism of blow-out fracture in a patient with extensive pneumatization of the sphenoid bone.

We report a patient with a lateral orbital wall fracture that occurred as a result of a blow-out fracture. The patient has extensive pneumatization of the sphenoid sinus, and the air extends into the lateral orbital wall. It makes the lateral orbital wall much thinner unusually, being more vulnerable to a sudden increase in intraorbital pressure. Pure blow-out fracture of the lateral orbital wall has not been reported in the medical literature. Therefore, this is the first case report of a lateral orbital wall fracture occurring as a blow-out mechanism.

[Primer on Generative Artificial Intelligence and Large Language Models in Medical Imaging]. / 의료영상에서 생성형 인공지능과 대형 언어 모델 ìž ë¬¸.

The recent advent of large language models (LLMs), such as ChatGPT, has drawn attention to generative artificial intelligence (AI) in a number of fields. Generative AI can produce different types of data including text, images, and voice, depending on the training methods and datasets used. Additionally, recent advancements in multimodal techniques, which can simultaneously process multiple data types like text and images, have expanded the potential of using multimodal generative AI in the medical environment where various types of clinical and imaging information are used together. This review summarizes the concepts and types of LLMs, image generative AI, and multimodal AI, and it examines the status and future possibilities of generative AI in the field of radiology.

Generative Adversarial Network with Robust Discriminator Through Multi-Task Learning for Low-Dose CT Denoising.

Reducing the dose of radiation in computed tomography (CT) is vital to decreasing secondary cancer risk. However, the use of low-dose CT (LDCT) images is accompanied by increased noise that can negatively impact diagnoses. Although numerous deep learning algorithms have been developed for LDCT denoising, several challenges persist, including the visual incongruence experienced by radiologists, unsatisfactory performances across various metrics, and insufficient exploration of the networks' robustness in other CT domains. To address such issues, this study proposes three novel accretions. First, we propose a generative adversarial network (GAN) with a robust discriminator through multi-task learning that simultaneously performs three vision tasks: restoration, image-level, and pixel-level decisions. The more multi-tasks that are performed, the better the denoising performance of the generator, which means multi-task learning enables the discriminator to provide more meaningful feedback to the generator. Second, two regulatory mechanisms, restoration consistency (RC) and non-difference suppression (NDS), are introduced to improve the discriminator's representation capabilities. These mechanisms eliminate irrelevant regions and compare the discriminator's results from the input and restoration, thus facilitating effective GAN training. Lastly, we incorporate residual fast Fourier transforms with convolution (Res-FFT-Conv) blocks into the generator to utilize both frequency and spatial representations. This approach provides mixed receptive fields by using spatial (or local), spectral (or global), and residual connections. Our model was evaluated using various pixel- and feature-space metrics in two denoising tasks. Additionally, we conducted visual scoring with radiologists. The results indicate superior performance in both quantitative and qualitative measures compared to state-of-the-art denoising techniques.

Supervised representation learning based on various levels of pediatric radiographic views for transfer learning.

Transfer learning plays a pivotal role in addressing the paucity of data, expediting training processes, and enhancing model performance. Nonetheless, the prevailing practice of transfer learning predominantly relies on pre-trained models designed for the natural image domain, which may not be well-suited for the medical image domain in grayscale. Recognizing the significance of leveraging transfer learning in medical research, we undertook the construction of class-balanced pediatric radiograph datasets collectively referred to as PedXnets, grounded in radiographic views using the pediatric radiographs collected over 24 years at Asan Medical Center. For PedXnets pre-training, approximately 70,000 X-ray images were utilized. Three different pre-training weights of PedXnet were constructed using Inception V3 for various radiation perspective classifications: Model-PedXnet-7C, Model-PedXnet-30C, and Model-PedXnet-68C. We validated the transferability and positive effects of transfer learning of PedXnets through pediatric downstream tasks including fracture classification and bone age assessment (BAA). The evaluation of transfer learning effects through classification and regression metrics showed superior performance of Model-PedXnets in quantitative assessments. Additionally, visual analyses confirmed that the Model-PedXnets were more focused on meaningful regions of interest.

Screening Patient Misidentification Errors Using a Deep Learning Model of Chest Radiography: A Seven Reader Study.

We aimed to evaluate the ability of deep learning (DL) models to identify patients from a paired chest radiograph (CXR) and compare their performance with that of human experts. In this retrospective study, patient identification DL models were developed using 240,004 CXRs. The models were validated using multiple datasets, namely, internal validation, CheXpert, and Chest ImaGenome (CIG), which include different populations. Model performance was analyzed in terms of disease change status. The performance of the models to identify patients from paired CXRs was compared with three junior radiology residents (group I), two senior radiology residents (group II), and two board-certified expert radiologists (group III). For the reader study, 240 patients (age, 56.617 ± 13.690 years, 113 females, 160 same pairs) were evaluated. A one-sided non-inferiority test was performed with a one-sided margin of 0.05. SimChest, our similarity-based DL model, demonstrated the best patient identification performance across multiple datasets, regardless of disease change status (internal validation [area under the receiver operating characteristic curve range: 0.992-0.999], CheXpert [0.933-0.948], and CIG [0.949-0.951]). The radiologists identified patients from the paired CXRs with a mean accuracy of 0.900 (95% confidence interval: 0.852-0.948), with performance increasing with experience (mean accuracy:group I [0.874], group II [0.904], group III [0.935], and SimChest [0.904]). SimChest achieved non-inferior performance compared to the radiologists (P for non-inferiority: 0.015). The findings of this diagnostic study indicate that DL models can screen for patient misidentification using a pair of CXRs non-inferiorly to human experts.

Approximating Intermediate Feature Maps of Self-Supervised Convolution Neural Network to Learn Hard Positive Representations in Chest Radiography.

Recent advances in contrastive learning have significantly improved the performance of deep learning models. In contrastive learning of medical images, dealing with positive representation is sometimes difficult because some strong augmentation techniques can disrupt contrastive learning owing to the subtle differences between other standardized CXRs compared to augmented positive pairs; therefore, additional efforts are required. In this study, we propose intermediate feature approximation (IFA) loss, which improves the performance of contrastive convolutional neural networks by focusing more on positive representations of CXRs without additional augmentations. The IFA loss encourages the feature maps of a query image and its positive pair to resemble each other by maximizing the cosine similarity between the intermediate feature outputs of the original data and the positive pairs. Therefore, we used the InfoNCE loss, which is commonly used loss to address negative representations, and the IFA loss, which addresses positive representations, together to improve the contrastive network. We evaluated the performance of the network using various downstream tasks, including classification, object detection, and a generative adversarial network (GAN) inversion task. The downstream task results demonstrated that IFA loss can improve the performance of effectively overcoming data imbalance and data scarcity; furthermore, it can serve as a perceptual loss encoder for GAN inversion. In addition, we have made our model publicly available to facilitate access and encourage further research and collaboration in the field.

Updated Primer on Generative Artificial Intelligence and Large Language Models in Medical Imaging for Medical Professionals.

The emergence of Chat Generative Pre-trained Transformer (ChatGPT), a chatbot developed by OpenAI, has garnered interest in the application of generative artificial intelligence (AI) models in the medical field. This review summarizes different generative AI models and their potential applications in the field of medicine and explores the evolving landscape of Generative Adversarial Networks and diffusion models since the introduction of generative AI models. These models have made valuable contributions to the field of radiology. Furthermore, this review also explores the significance of synthetic data in addressing privacy concerns and augmenting data diversity and quality within the medical domain, in addition to emphasizing the role of inversion in the investigation of generative models and outlining an approach to replicate this process. We provide an overview of Large Language Models, such as GPTs and bidirectional encoder representations (BERTs), that focus on prominent representatives and discuss recent initiatives involving language-vision models in radiology, including innovative large language and vision assistant for biomedicine (LLaVa-Med), to illustrate their practical application. This comprehensive review offers insights into the wide-ranging applications of generative AI models in clinical research and emphasizes their transformative potential.

Screening of adolescent idiopathic scoliosis using generative adversarial network (GAN) inversion method in chest radiographs.

OBJECTIVE: Conventional computer-aided diagnosis using convolutional neural networks (CNN) has limitations in detecting sensitive changes and determining accurate decision boundaries in spectral and structural diseases such as scoliosis. We devised a new method to detect and diagnose adolescent idiopathic scoliosis in chest X-rays (CXRs) employing the latent space's discriminative ability in the generative adversarial network (GAN) and a simple multi-layer perceptron (MLP) to screen adolescent idiopathic scoliosis CXRs. MATERIALS AND METHODS: Our model was trained and validated in a two-step manner. First, we trained a GAN using CXRs with various scoliosis severities and utilized the trained network as a feature extractor using the GAN inversion method. Second, we classified each vector from the latent space using a simple MLP. RESULTS: The 2-layer MLP exhibited the best classification in the ablation study. With this model, the area under the receiver operating characteristic (AUROC) curves were 0.850 in the internal and 0.847 in the external datasets. Furthermore, when the sensitivity was fixed at 0.9, the model's specificity was 0.697 in the internal and 0.646 in the external datasets. CONCLUSION: We developed a classifier for Adolescent idiopathic scoliosis (AIS) through generative representation learning. Our model shows good AUROC under screening chest radiographs in both the internal and external datasets. Our model has learned the spectral severity of AIS, enabling it to generate normal images even when trained solely on scoliosis radiographs.

Overcoming the Challenges in the Development and Implementation of Artificial Intelligence in Radiology: A Comprehensive Review of Solutions Beyond Supervised Learning.

Artificial intelligence (AI) in radiology is a rapidly developing field with several prospective clinical studies demonstrating its benefits in clinical practice. In 2022, the Korean Society of Radiology held a forum to discuss the challenges and drawbacks in AI development and implementation. Various barriers hinder the successful application and widespread adoption of AI in radiology, such as limited annotated data, data privacy and security, data heterogeneity, imbalanced data, model interpretability, overfitting, and integration with clinical workflows. In this review, some of the various possible solutions to these challenges are presented and discussed; these include training with longitudinal and multimodal datasets, dense training with multitask learning and multimodal learning, self-supervised contrastive learning, various image modifications and syntheses using generative models, explainable AI, causal learning, federated learning with large data models, and digital twins.

Bone suppression on pediatric chest radiographs via a deep learning-based cascade model.

BACKGROUND AND OBJECTIVE: Bone suppression images (BSIs) of chest radiographs (CXRs) have been proven to improve diagnosis of pulmonary diseases. To acquire BSIs, dual-energy subtraction (DES) or a deep-learning-based model trained with DES-based BSIs have been used. However, neither technique could be applied to pediatric patients owing to the harmful effects of DES. In this study, we developed a novel method for bone suppression in pediatric CXRs. METHODS: First, a model using digitally reconstructed radiographs (DRRs) of adults, which were used to generate pseudo-CXRs from computed tomography images, was developed by training a 2-channel contrastive-unpaired-image-translation network. Second, this model was applied to 129 pediatric DRRs to generate the paired training data of pseudo-pediatric CXRs. Finally, by training a U-Net with these paired data, a bone suppression model for pediatric CXRs was developed. RESULTS: The evaluation metrics were peak signal to noise ratio, root mean absolute error and structural similarity index measure at soft-tissue and bone region of the lung. In addition, an expert radiologist scored the effectiveness of BSIs on a scale of 1-5. The obtained result of 3.31 ± 0.48 indicates that the BSIs show homogeneous bone removal despite subtle residual bone shadow. CONCLUSION: Our method shows that the pixel intensity at soft-tissue regions was preserved, and bones were well subtracted; this can be useful for detecting early pulmonary disease in pediatric CXRs.