The Ewha Medical Journal

[English]

Comparative evaluation of deep learning architectures, including UNet, TransUNet, and MIST, for left atrium segmentation in cardiac computed tomography of congenital heart diseases: Seoyeong Yun, Jooyoung Choi; Received February 26, 2025 Accepted April 10, 2025 Published online April 21, 2025; DOI: https://doi.org/10.12771/emj.2025.00087 [Epub ahead of print]

Purpose
This study compares 3 deep learning models (UNet, TransUNet, and MIST) for left atrium (LA) segmentation of cardiac computed tomography (CT) images from patients with congenital heart disease (CHD). It investigates how architectural variations in the MIST model, such as spatial squeeze-and-excitation attention, impact Dice score and HD95.
Methods
We analyzed 108 publicly available, de-identified CT volumes from the ImageCHD dataset. Volumes underwent resampling, intensity normalization, and data augmentation. UNet, TransUNet, and MIST models were trained using 80% of 97 cases, with the remaining 20% employed for validation. Eleven cases were reserved for testing. Performance was evaluated using the Dice score (measuring overlap accuracy) and HD95 (reflecting boundary accuracy). Statistical comparisons were performed via one-way repeated measures analysis of variance.
Results
MIST achieved the highest mean Dice score (0.74; 95% confidence interval, 0.67–0.81), significantly outperforming TransUNet (0.53; P<0.001) and UNet (0.49; P<0.001). Regarding HD95, TransUNet (9.09 mm) and MIST (5.77 mm) similarly outperformed UNet (27.49 mm; P<0.0001). In ablation experiments, the inclusion of spatial attention did not further enhance the MIST model’s performance, suggesting redundancy with existing attention mechanisms. However, the integration of multi-scale features and refined skip connections consistently improved segmentation accuracy and boundary delineation.
Conclusion
MIST demonstrated superior LA segmentation, highlighting the benefits of its integrated multi-scale features and optimized architecture. Nevertheless, its computational overhead complicates practical clinical deployment. Our findings underscore the value of advanced hybrid models in cardiac imaging, providing improved reliability for CHD evaluation. Future studies should balance segmentation accuracy with feasible clinical implementation.

68 View
8 Download

[English]

Feature-based ensemble modeling for addressing diabetes data imbalance using the SMOTE, RUS, and random forest methods: a prediction study: Younseo Jang; Received March 31, 2025 Accepted April 10, 2025 Published online April 15, 2025; DOI: https://doi.org/10.12771/emj.2025.00353 [Epub ahead of print]

Purpose
This study developed and evaluated a feature-based ensemble model integrating the synthetic minority oversampling technique (SMOTE) and random undersampling (RUS) methods with a random forest approach to address class imbalance in machine learning for early diabetes detection, aiming to improve predictive performance.
Methods
Using the Scikit-learn diabetes dataset (442 samples, 10 features), we binarized the target variable (diabetes progression) at the 75th percentile and split it 80:20 using stratified sampling. The training set was balanced to a 1:2 minority-to-majority ratio via SMOTE (0.6) and RUS (0.66). A feature-based ensemble model was constructed by training random forest classifiers on 10 two-feature subsets, selected based on feature importance, and combining their outputs using soft voting. Performance was compared against 13 baseline models, using accuracy and area under the curve (AUC) as metrics on the imbalanced test set.
Results
The feature-based ensemble model and balanced random forest both achieved the highest accuracy (0.8764), followed by the fully connected neural network (0.8700). The ensemble model had an excellent AUC (0.9227), while k-nearest neighbors had the lowest accuracy (0.8427). Visualizations confirmed its superior discriminative ability, especially for the minority (high-risk) class, which is a critical factor in medical contexts.
Conclusion
Integrating SMOTE, RUS, and feature-based ensemble learning improved classification performance in imbalanced diabetes datasets by delivering robust accuracy and high recall for the minority class. This approach outperforms traditional resampling techniques and deep learning models, offering a scalable and interpretable solution for early diabetes prediction and potentially other medical applications.

151 View
9 Download

[English]

Cyclic dual latent discovery for improved blood glucose prediction through patient–provider interaction modeling: a prediction study: Suyeon Park, Seoyoung Kim, Dohyoung Rim; Received March 30, 2025 Accepted April 7, 2025 Published online April 15, 2025; DOI: https://doi.org/10.12771/emj.2025.00332 [Epub ahead of print]

Purpose
Accurate prediction of blood glucose variability is crucial for effective diabetes management, as both hypoglycemia and hyperglycemia are associated with increased morbidity and mortality. However, conventional predictive models rely primarily on patient-specific biometric data, often neglecting the influence of patient–provider interactions, which can significantly impact outcomes. This study introduces Cyclic Dual Latent Discovery (CDLD), a deep learning framework that explicitly models patient–provider interactions to improve prediction of blood glucose levels. By leveraging a real-world intensive care unit (ICU) dataset, the model captures latent attributes of both patients and providers, thus improving forecasting accuracy.
Methods
ICU patient records were obtained from the MIMIC-IV v3.0 critical care database, including approximately 5,014 instances of patient–provider interaction. The CDLD model uses a cyclic training mechanism that alternately updates patient and provider latent representations to optimize predictive performance. During preprocessing, all numeric features were normalized, and extreme glucose values were capped at 500 mg/dL to mitigate the effect of outliers.
Results
CDLD outperformed conventional models, achieving a root mean square error of 0.0852 on the validation set and 0.0899 on the test set, which indicates improved generalization. The model effectively captured latent patient–provider interaction patterns, yielding more accurate glucose variability predictions than baseline approaches.
Conclusion
Integrating patient–provider interaction modeling into predictive frameworks can increase blood glucose prediction accuracy. The CDLD model offers a novel approach to diabetes management, potentially paving the way for artificial intelligence-driven personalized treatment strategies.

157 View
7 Download

Special topic: recent clinical approach to shoulder diseases in older adults

[English]

Application of deep learning for diagnosis of shoulder diseases in older adults: a narrative review: Sung Min Rhee; Ewha Med J 2025;48(1):e6. Published online January 31, 2025; DOI: https://doi.org/10.12771/emj.2025.e6

Abstract
PDF

Shoulder diseases pose a significant health challenge for older adults, often causing pain, functional decline, and decreased independence. This narrative review explores how deep learning (DL) can address diagnostic challenges by automating tasks such as image segmentation, disease detection, and motion analysis. Recent research highlights the effectiveness of DL-based convolutional neural networks and machine learning frameworks in diagnosing various shoulder pathologies. Automated image analysis facilitates the accurate assessment of rotator cuff tear size, muscle degeneration, and fatty infiltration in MRI or CT scans, frequently matching or surpassing the accuracy of human experts. Convolutional neural network-based systems are also adept at classifying fractures and joint conditions, enabling the rapid identification of common causes of shoulder pain from plain radiographs. Furthermore, advanced techniques like pose estimation provide precise measurements of the shoulder joint's range of motion and support personalized rehabilitation plans. These automated approaches have also been successful in quantifying local osteoporosis, utilizing machine learning-derived indices to classify bone density status. DL has demonstrated significant potential to improve diagnostic accuracy, efficiency, and consistency in the management of shoulder diseases in older patients. Machine learning-based assessments of imaging data and motion parameters can help clinicians optimize treatment plans and improve patient outcomes. However, to ensure their generalizability, reproducibility, and effective integration into routine clinical workflows, large-scale, prospective validation studies are necessary. As data availability and computational resources increase, the ongoing development of DL-driven applications is expected to further advance and personalize musculoskeletal care, benefiting both healthcare providers and the aging population.

223 View
7 Download

Original Articles

[English]

An accurate pediatric bone age prediction model using deep learning and contrast conversion: Dong Hyeok Choi, So Hyun Ahn, Rena Lee; Ewha Med J 2024;47(2):e23. Published online April 30, 2024; DOI: https://doi.org/10.12771/emj.2024.e23

Abstract
PDF

Objectives:

This study aimed to develop an accurate pediatric bone age prediction model by utilizing deep learning models and contrast conversion techniques, in order to improve growth assessment and clinical decision-making in clinical practice.

Methods:

The study employed a variety of deep learning models and contrast conversion techniques to predict bone age. The training dataset consisted of pediatric left-hand X-ray images, each annotated with bone age and sex information. Deep learning models, including a convolutional neural network , Residual Network 50 , Visual Geometry Group 19, Inception V3, and Xception were trained and assessed using the mean absolute error (MAE). For the test data, contrast conversion techniques including fuzzy contrast enhancement, contrast limited adaptive histogram equalization (HE) , and HE were implemented. The quality of the images was evaluated using peak signal-to-noise ratio (SNR), mean squared error, SNR, coefficient of variation, and contrast-to-noise ratio metrics. The bone age prediction results using the test data were evaluated based on the MAE and root mean square error, and the t-test was performed.

Results:

The Xception model showed the best performance (MAE=41.12). HE exhibited superior image quality, with higher SNR and coefficient of variation values than other methods. Additionally, HE demonstrated the highest contrast among the techniques assessed, with a contrast-to-noise ratio value of 1.29. Improvements in bone age prediction resulted in a decline in MAE from 2.11 to 0.24, along with a decrease in root mean square error from 0.21 to 0.02.

Conclusion:

This study demonstrates that preprocessing the data before model training does not significantly affect the performance of bone age prediction when comparing contrast-converted images with original images.

Citations

Citations to this article as recorded by

Gender equity in medicine, artificial intelligence, and other articles in this issue
Sun Huh
The Ewha Medical Journal.2024;[Epub] CrossRef

200 View
4 Download
1 Crossref

[English]

What is the role of artificial intelligence in general surgery?: Seung Min Baik, Ryung-Ah Lee; Ewha Med J 2024;47(2):e22. Published online April 30, 2024; DOI: https://doi.org/10.12771/emj.2024.e22

Abstract
PDF

The capabilities of artificial intelligence (AI) have recently surged, largely due to advancements in deep learning inspired by the structure and function of the neural networks of the human brain. In the medical field, the impact of AI spans from diagnostics and treatment recommendations to patient engagement and monitoring, considerably improving efficiency and outcomes. The clinical integration of AI has also been examined in specialties, including pathology, radiology, and oncology. General surgery primarily involves manual manipulation and includes preoperative, intraoperative, and postoperative care, all of which are critical for saving lives. Other fields have strived to utilize and adopt AI; nonetheless, general surgery appears to have retrogressed. In this review, we analyzed the published research, to understand how the application of AI in general surgery differs from that in other medical fields. Based on previous research in other fields, the application of AI in the preoperative stage is nearing feasibility. Ongoing research efforts aim to utilize AI to improve and predict operative outcomes, enhance performance, and improve patient care. However, the use of AI in the operating room remains significantly understudied. Moreover, ethical responsibilities are associated with such research, necessitating extensive work to gather evidence. By fostering interdisciplinary collaboration and leveraging lessons from AI success stories in other fields, AI tools could be specifically tailored for general surgery. Surgeons should be prepared for the integration of AI into clinical practice to achieve better outcomes; therefore, the time has come to consider ethical and legal implications.

Citations

Citations to this article as recorded by