Ji-hun BAE

Mr. Ji-hun BAE

orcid

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
7 July 2025 in IEEE Transactions on Network Science and Engineering
Verified
Journal Article
Open access
HLHunmin LEE
Hunmin LEE
HSHongju Seong
Hongju Seong
JBJi-hun BAE
Ji-hun BAE
WKWonbin Kim
Wonbin Kim
HKHyeokchan Kwon
Hyeokchan Kwon
Daehee SeoDaehee Seo
Daehee Seo
Federated Learning (FL)-based modulation decoding enables collaborative learning across distributed Artificial Intelligence (AI)-powered modulational signal decoding models in large-scale wireless communication networks. Despite significant advancements, FL-based modulation decoding faces two primary challenges: (1) Vulnerability to noisy environments - Current studies often prioritize high-quality datasets during training to ensure model performance, which often results in overfitting since real-world signals involve a wide range of noise levels from various sources. (2) Model divergence in non-Independent and Identically Distributed (non-IID) environments - The aggregation step in a conventional FL leads to divergence on a large federated scale, resulting in the loss of local-specific knowledge. Moreover, existing FL-based modulation decoding studies typically focus on a single non-IID issue, which limits their ability to generalize across large-scale heterogeneous settings. To address these challenges, we propose FedVaccine, a framework that enhances model generalization across diverse noise conditions and non-IID data environments. Instead of relying on high-quality datasets that risk overfitting, FedVaccine employs a harmonic noise resilience approach—akin to a vaccination effect—that deliberately exposes the model to controlled noise, identifying noise tolerance, regularizing training, and ensuring robust decoding performance across varying noise levels. Additionally, FedVaccine incorporates a Federated Split Learning approach to cumulatively update the global model, ensuring adaptive recalibration and mitigating the loss of local knowledge during aggregation. Furthermore, a localized support set is used to ensure effective personalization. Experimental results demonstrate that FedVaccine outperforms existing FL-based modulation classification models and other learning paradigms in both IID and non-IID scenarios using public datasets, showcasing its effectiveness in practical deployment scenarios across wireless networks.

31 December 2024 in Biomedical Physics & Engineering Express
Verified
Journal Article
Open access
JBJi-hun BAE
Ji-hun BAE
HLHunmin LEE
Hunmin LEE
JHJinglu Hu
Jinglu Hu
Recent studies on graph representation learning in brain tumor learning tasks have garnered significant interest by encoding and learning inherent relationships among the geometric features of tumors. A prominent challenge in brain tumor MRI datasets is the issue of class imbalance. While deep learning architectures such as CNN- and Transformer-based models have shown promise in addressing this challenge through their complex models with large parameters spaces., Graph-based networks face significant difficulties. Specifically, graph networks often struggle with issues like over-smoothing and oscillation convergence, limiting their effectiveness in such tasks. To address these challenges, we propose novel graph spectral convolutional networks called HeatGSNs, which incorporate eigenfilters and learnable low-pass graph heat kernels to capture geometric similarities within tumor classes. HeatGSNs utilize a continuous feature propagation mechanism derived from the forward finite difference of graph heat kernels, which is approximated by the cosine form of the shift-scaled Chebyshev polynomial and modified Bessel functions, ensuring efficient and accurate performance achievement. Experimental results on the BRATS2021 dataset demonstrate the effectiveness of HeatGSNs, achieving an average Dice score of 90%, a Hausdorff Distance (95%) of 5.45 mm, and an average accuracy of 80.11%. Furthermore, HeatGSNs require significantly fewer parameters than existing methods, with an average of 1.79M parameters, highlighting efficiency and effectiveness.

10 September 2025 in Electronics
Verified
Journal Article
Open access
PDF file
JLJin Lee
Jin Lee
JBJi-hun BAE
Ji-hun BAE
DVDang Thanh Vu
Dang Thanh Vu
ALAnh Le
Anh Le
ZRZahid Ur Rahman
Zahid Ur Rahman
JKJin Young Kim
Jin Young Kim
Foundation models, exemplified by the Segment Anything Model (SAM), have revolutionized object segmentation with their impressive zero-shot capabilities. The recent SAM2 extended these abilities to the video domain, utilizing an object pointer and memory attention to maintain temporal segment consistency. However, a critical limitation of SAM2 is its vulnerability to error accumulation, where an initial incorrect mask can propagate through subsequent frames, leading to tracking failure. To address this, we propose a novel method that actively monitors the temporal segment consistency of masks by evaluating the distance of object pointers across frames. When a potential error is detected via a sharp increase in distance, our method triggers a particle filter based re-inference module. This framework models object’s motion to predict a corrected bounding box, effectively guiding the model to recover the valid mask and preventing error propagation. Extensive zero-shot evaluations on DAVIS, LVOS v2, YouTube-VOS and qualitative results show that the proposed, parameter-free procedure consistently improves temporal coherence, raising mean IoU by 0.1 on DAVIS, by 0.13 on the LVOS v2 train split and 0.05 on the LVOS v2 validation split, and by 0.02 on YouTube-VOS, thereby offering a simple and effective route to more robust video object segmentation with SAM2.

24 August 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
JBJi-hun BAE
Ji-hun BAE
YBYeonho Bae
Yeonho Bae
JHJinglu Hu
Jinglu Hu
Deciding whether a vertex of a cubical complex lies on a locally well-composed surface reduces to a finite question: examine the occupancy pattern of the eight voxels incident to the vertex and decide whether the resulting active incidence structure is connected and 2-regular. We give a symbolic account of this decision. Writing \(n_E(S)\) and \(n_F(S)\) for the numbers of active edges and faces determined by elementary local activity rules, we show that every active face has degree exactly two and every active edge has degree two or four, yielding the identity \(n_F(S)=n_E(S)+k(S)\) for a single defect count \(k(S)\), and hence a surface criterion equivalent to \(\mathrm{comps}(S)=1\wedge n_E(S)=n_F(S)\). We further prove that the active incidence structure is invariant under complementation as a literal identity, not merely up to isomorphism, and combine this duality with a finite case analysis to classify explicitly which pairs \((|S|,d_1(S))\) of cardinality and Hamming-adjacency count are realized by a surface state, yielding an exact threshold classification as a proved corollary. The classification is established by a complete finite case analysis up to cube symmetries for \(|S|\le4\) and by duality for \(|S|\in\{5,6,7\}\).

28 January 2026 in Biomedical Physics & Engineering Express
Verified
Journal Article
Closed access
JBJi-hun BAE
Ji-hun BAE
HLHunmin LEE
Hunmin LEE
JHJinglu Hu
Jinglu Hu
Segmenting glioblastoma in medical imaging remains challenging due to the tumor’s irregular shape, heterogeneous texture, and poorly defined boundaries, which often lead to inaccurate delineation by conventional methods. To address these challenges, we propose a novel segmentation framework that leverages Topological Data Analysis (TDA) to capture intrinsic topological features of gliomas, reducing reliance on large annotated datasets while improving structural fidelity. Our approach exploits TDA’s interpretable filtrations and persistent homology alongside explicit spatial adjacency information to enhance segmentation accuracy. The method proceeds sequentially: (1) Whole Tumor (WT) segmentation via an Automated Thresholding Algorithm on reverse filtration (R-filtration), (2) Enhancing Tumor (ET) segmentation using carefully selected persistent homology features, and (3) non-enhancing Tumor Core (TC) and Edema (ED) segmentation based on their spatial adjacency with ET through a criterion-driven iterative process. To quantify segmentation performance, we introduce a novel fuzzy Edge-Dice score applicable across all steps. Evaluated on the public BRATS2021 and BRATS2022-Reg datasets, our TDA-based framework achieves robust and accurate segmentation, highlighting the potential of TDA methods to complement or surpass conventional deep learning approaches in real-world medical imaging applications.

19 August 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
JBJi-hun BAE
Ji-hun BAE
YBYeonho Bae
Yeonho Bae
JHJinglu Hu
Jinglu Hu
Local topological singularities on digital or voxel grids motivated well-composedness conditions excluding them; P-well-composedness lifts this regularity into an intrinsic, order-theoretic setting on posets. We study whether P-well-composedness is inherited by the strict neighborhood of a face in a finite embedded cubical complex. Let X = F(K) be the face poset of a finite nonempty set of grid n-cubes, with ambient rank n ≤ 3, and let Nh = θ′X (h) carry the induced order. We prove Nh is P-well-composed whenever X is, provided h is a vertex or has no strict coface in the border ΔX, via a coherence argument for vertices and a finite, computer-assisted enumeration otherwise. The remaining rank-three case, an edge with a strict coface in the border, is posed as a No-Shrinkage Conjecture, with computational evidence reported separately.