https://scholar.dgist.ac.kr/handle/20.500.11750/1923 2026-04-06T10:01:11Z 2026-04-06T10:01:11Z Kim, Jihyeon Lee, Gyeongmin Shin, Seung Yeon Kim, Soopil Park, Sang Hyun https://scholar.dgist.ac.kr/handle/20.500.11750/59400 2026-01-21T12:10:15Z 2025-11-30T15:00:00Z

Title: MOSInversion: Knowledge distillation-based incremental learning in organ segmentation using DeepInversion Author(s): Kim, Jihyeon; Lee, Gyeongmin; Shin, Seung Yeon; Kim, Soopil; Park, Sang Hyun Abstract: Despite recent advancements in multi-organ segmentation (MOS) of medical images, existing models are limited in terms of extending their capability to unseen classes. Incremental learning has been proposed to enable models to learn new classes progressively, possibly using multiple datasets from different institutions. In this setting, models easily experience performance degradation on previously learned classes i.e., catastrophic forgetting. Although many methods have been proposed to mitigate this issue, applying them to medical imaging applications like multi-organ segmentation is not easy due to the large memory requirement when used for 3D medical data such as CT scans or the need for additional training of a generator for image synthesis. In this paper, we propose an incremental learning framework that leverages diverse synthetic images to retain the knowledge learned from previously seen data. We design MOSInversion to generate the synthetic images by utilizing a pre-trained model from the previous step. MOSInversion generates diverse images by using segmentation masks so that we can manipulate the shape, location, and size of organs. We evaluate our proposed method using three abdominal CT datasets (FLARE21, MSD, and KiTS19) and achieve state-of-the-art accuracy.

2025-11-30T15:00:00Z Lee, Gyeongmin 박무열 서연우 소정태 정영준 Mikiko Ito 이병기 박상현 https://scholar.dgist.ac.kr/handle/20.500.11750/59370 2026-01-15T13:10:10Z 2025-10-31T15:00:00Z

Title: 해부학적 제약과 교차 어텐션을 통한 CT 및 디지털 단층촬영술의 자가 지도 비강체 정합 Author(s): Lee, Gyeongmin; 박무열; 서연우; 소정태; 정영준; Mikiko Ito; 이병기; 박상현 Abstract: In image-guided surgery, the registration of preoperative 3D computed tomography (CT) and intraoperative digital tomosynthesis (DTS) images is essential. However, it presents significant technical challenges due to the multi-modal nature of the two images, inherent DTS artifacts, and the lack of ground truth data. Therefore, this study proposes a self-supervised learning-based non-rigid registration framework. The proposed method precisely estimates local deformations through deep learning-based non-rigid registration, leveraging pre-registration on CT–DTS image pairs. To overcome the lack of ground truth data, a training data pipeline was established. This pipeline generates CT-synthesized DTS-ground truth deformation field data pairs by applying anatomically constrained virtual deformations to the CT images and re-projecting them. Additionally, we designed a specialized network architecture incorporating a multi-encoder and a cross-attention mechanism to effectively fuse the features of the multi-modal images. Experimental results using a public dataset show that the proposed method achieved a 3D target registration error of 12.99 mm. This study is expected to contribute to the future advancement of surgical navigation systems by offering a new direction for the CT–DTS registration problem.

2025-10-31T15:00:00Z An, Sion Kang, Myeongkyun Kim, Soopil Chikontwe, Philip Shen, Li Park, Sang Hyun https://scholar.dgist.ac.kr/handle/20.500.11750/58900 2025-08-13T08:40:12Z 2025-12-31T15:00:00Z

Title: Subject-adaptive meta-learning for personalized BCI: A fusion of resting-state EEG signal and task-specific information Author(s): An, Sion; Kang, Myeongkyun; Kim, Soopil; Chikontwe, Philip; Shen, Li; Park, Sang Hyun Abstract: Electroencephalography (EEG) motor imagery (MI) classification is fundamental to understanding the neural mechanisms underlying human movement and advancing brain-computer interfaces (BCI) applications. Deep learning based approaches have demonstrated exceptional proficiency in classifying EEG signals. However, their applications are often restricted by the large variation of signals between individuals, i.e., inter-subject variability. To mitigate this issue, some studies have employed task-specific (TS) EEG signals recorded from the target subject, thereby improving classification performance. Despite this progress, collecting TS EEG data remains a major limitation due to its time-consuming and labor-intensive process. Conversely, resting state (RS) EEG signals present a promising alternative, as they can be acquired more easily and contain rich subject information. In this paper, we propose a subject-adaptive learning approach using RS EEG signals within a meta-learning framework. The model learns to adapt to each subject using only their RS EEG signals for personalized EEG MI classification. Our learning framework consists of two iterative phases. In the subject-specific training phase, we fuse RS EEG signals with TS information while retaining individual subject characteristics and use the fused signals to adapt the model to the target subject. In the meta-training phase, the model predicts the MI class corresponding to the given TS EEG signals and computes the loss to update the meta-parameters for rapid target adaptation. Our method achieves an average accuracy improvement of 10.05% across two encoders and three benchmark datasets. Furthermore, visualization results show that the fused RS EEG signals combined with TS information exhibit characteristics similar to real TS EEG signals. These findings highlight the potential of leveraging RS EEG signals to advance practical BCI systems.

2025-12-31T15:00:00Z Kang, Myeongkyun Kim, Soopil Jin, Kyong Hwan Adeli, Ehsan Pohl, Kilian M. Park, Sang Hyun https://scholar.dgist.ac.kr/handle/20.500.11750/58643 2025-07-25T02:47:47Z 2024-04-30T15:00:00Z

Title: FedNN: Federated learning on concept drift data using weight and adaptive group normalizations Author(s): Kang, Myeongkyun; Kim, Soopil; Jin, Kyong Hwan; Adeli, Ehsan; Pohl, Kilian M.; Park, Sang Hyun Abstract: Federated Learning (FL) allows a global model to be trained without sharing private raw data. The major challenge in FL is client-wise data heterogeneity leading to different model convergence speed and accuracy. Despite the recent progress of FL, most methods verify their accuracy on prior probability shift (label distribution skew) dataset, while the concept drift problem (i.e., where each client has distinct styles of input while sharing the same labels) has not been explored. In real scenarios, concept drift is of paramount concern in FL since the client's data is collected under extremely different conditions making FL optimization more challenging. Significant differences in inputs among clients exacerbate the heterogeneity of clients’ parameters compared to prior probability shift, ultimately resulting in failures for previous FL approaches. To address the challenge of concept drift, we use Weight Normalization (WN) and Adaptive Group Normalization (AGN) to alleviate conflicts during global model updates. WN re-parameterizes weights to have zero mean and unit variance while AGN adaptively selects the optimal mean and standard deviation for feature normalization based on the dataset. These two components significantly contribute to having consistent activations after global model updates reducing heterogeneity in concept drift data. Comprehensive experiments on seven datasets (with concept drift) demonstrate that our method outperforms five state-of-the-art FL methods and shows faster convergence speed compared to the previous FL methods. © 2024 Elsevier Ltd

2024-04-30T15:00:00Z