RESEARCH
Embedded Files
MEDICAL AI
ADSLab applies artificial intelligence algorithms to analyze diverse medical domain data, aiming to enhance diagnostic accuracy and healthcare solutions. Our research encompasses various types of medical data, including imaging data such as MRI, CT, and X-ray scans, structured and unstructured patient information from Electronic Medical Records (EMR) and Electronic Health Records (EHR), as well as time-series data collected from health monitoring sensors. We can access to extensive datasets from eight major hospitals within The Catholic University of Korea’s healthcare network: Seoul St. Mary’s Hospital, Yeouido St. Mary’s Hospital, Uijeongbu St. Mary’s Hospital, Bucheon St. Mary’s Hospital, Eunpyeong St. Mary’s Hospital, Incheon St. Mary’s Hospital, St. Vincent’s Hospital, and Daejeon St. Mary’s Hospital.
Research Detatils :
🔵Skin Lesion Classification : Skin cancer classification is a complex task due to the subtle visual differences between lesion types and the limited availability of labeled medical images. To address these challenges, ADSLab has developed two complementary approaches. The first is the ABC ensemble model, which incorporates dermatological domain knowledge—specifically asymmetry, border, and color—into the preprocessing and model ensemble process. The second is the AGS network, which improves data diversity and feature learning by integrating three modules: Augmentation, GAN-based synthetic image generation, and Segmentation. Both methods were evaluated on the HAM10000 dataset and achieved notable improvements in classification accuracy and interpretability compared to baseline models.
🔵Spinal Data Analysis : ADSLab analyzes multimodal spinal data to support diagnosis, prognosis, and treatment planning for spinal disorders. Our work integrates radiological imaging (MRI, CT), clinical records, and histopathological slides. We extract and quantify key features such as vertebral compression fractures (VCF), endplate morphology, and muscle-fat ratios using AI-based segmentation and measurement tools. Additionally, we perform statistical analysis on clinical variables and use machine learning models to predict outcomes such as malignancy, progression, or surgical need. Our models are validated on multi-institutional datasets, enhancing their generalizability and clinical relevance.
🔵Gastric Cancer Metastasis Prediction : ADSLab investigates predictive models for lymph node metastasis and distant spread in gastric cancer patients using structured clinical data and pathology images. We utilize feature engineering and machine learning algorithms to identify high-risk patients preoperatively. Our models aim to assist clinicians in making informed decisions regarding surgical extent and adjuvant therapies, ultimately improving patient outcomes.
🔵Brain Tumor Prediction : ADSLab is developing a brain tumor prediction model using the BraTS dataset, which includes multimodal 3D MRI scans of glioma patients. To reflect the importance of image slice order in 3D data, we combine vision-based methods with temporal modeling. Specifically, we explore hybrid models that integrate attention mechanisms and Mamba, enabling efficient capture of spatial-temporal patterns for improved segmentation and classification.
Time-series Analysis & Anomaly Detection
ADSLab conducts research on time series analysis and anomaly detection to extract meaningful patterns and detect irregularities in temporal data. We develop advanced machine learning and deep learning algorithms to model complex temporal dependencies, forecast future trends, and identify abnormal events across various domains. Our approaches include recurrent neural networks (RNNs), temporal convolutional networks (TCNs), transformers, and probabilistic models, as well as recent architectures such as Mamba, TimesFM, and PatchAD. These methods are applied to diverse time series datasets such as patient vital signs, satellite telemetry data, financial indicators, and industrial sensor streams. Key applications include anomaly detection in medical and industrial systems, orbit prediction for satellite trajectory forecasting, and prognosis prediction in clinical and biomedical contexts. By integrating forecasting and detection capabilities, we aim to enhance decision-making in critical time-dependent environments.
Research Details:
🔵 Orbit Prediction : We proposed the DASR (Decomposed Attention Segment Recurrent Neural Network) model, which integrates Multi-Head Attention and Tensor Train Decomposition into SegRNN for improved satellite orbit prediction. Using real-world datasets, DASR outperformed existing state-of-the-art models with significantly reduced prediction error and model size.
🔵 Tensorized attention : We applied Tensor Train (TT) Decomposition to the attention mechanism to reduce the number of parameters while preserving prediction accuracy. Specifically, the weight matrices of query, key, and value in Multi-Head Attention are stacked into a 3D tensor and decomposed into smaller core tensors using TT. This approach significantly compresses the attention weights and enables efficient modeling of high-dimensional time series data.
Other Research Areas
In addition to our core research domains, ADSLab engages in a variety of emerging topics that reflect the evolving landscape of AI and data science.
Research Details:
Blurred as the work is not yet published.
🔵Machine Unlearning :
We explore methods for selectively removing the influence of specific data from trained models. This research supports model flexibility and addresses growing concerns about data retention and user control.
Blurred as the work is not yet published.
🔵Biometric Authentication :
Our lab studies AI-driven techniques for identity verification using physiological features like fingerprints and facial images. These methods aim to improve the security and reliability of authentication systems.
🔵Data De-identification :
We investigate ways to anonymize sensitive personal and medical data, ensuring privacy protection while retaining the analytical value of the data for machine learning and research.
🔵Deepfake Detection :
As synthetic media becomes more prevalent, we develop models to detect deepfakes in images and videos. This work contributes to digital content integrity and the prevention of misinformation.
Page updated
Google Sites
Report abuse