Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data—the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.

Immune status critically affects cancer progression and therapy responses. This study aimed to identify plasma proteome changes in immunosuppressive cancer and potential biomarkers predicting systemic immunosuppression. Mouse models of syngeneic breast tumors (benign 67NR and malignant 4T1) were used to collect plasma samples. Plasma samples from naive mice and both early- and late-stage tumor-bearing mice were subjected to liquid chromatography-mass spectrometry (LC-MS) analysis. 4T1-bearing mice showed systemic immunosuppression characterized by significant generation of myeloid-derived suppressor cells (MDSCs) as early as 7 days after tumor implantation, unlike 67NR tumors. LC-MS identified 1086 proteins across the five experimental groups, with 27 proteins showing group-specific expression in 4T1 blood compared with 67NR blood. Immune-related proteins osteopontin, lactotransferrin, calreticulin, and peroxiredoxin 2 were selected as potential biomarkers of MDSC-producing breast cancer. These markers were expressed in cancer cells or MDSC in the 4T1 model, and osteopontin and peroxiredoxin 2 were associated with low survival probability and high recurrence in patients with triple-negative breast cancer. Our findings suggest that MDSC-producing immunosuppressive cancers have unique plasma proteomes, offering additional insights into cancer immune status.

Ethnopharmacological relevance: Concurrent use of traditional herbal medicines and conventional drugs, particularly for stroke treatment, is widespread, raising concerns about potential drug interactions. Aim of the study: This clinical study aimed to investigate interactions between edoxaban, a direct oral anticoagulant, and two traditional herbal medicines commonly used for stroke: Banhabaekchulcheonmatang (BBCT) and Hwangryeonhaedoktang (HRHDT). Materials and methods: Korean healthy volunteers participated in a randomized, open-label, three-period, three- treatment, two-sequence clinical study. Treatments consisted of a single oral dose of edoxaban tablet (60 mg) in the presence or absence of multiple doses of BBCT or HRHDT three times daily for six days. Pharmacokinetic and pharmacodynamic parameters of edoxaban and its active metabolite M4 were assessed following administration of edoxaban alone or in co-administration with BBCT or HRHDT. Results: When edoxaban was co-administered with BBCT or HRHDT, the area under the curve (AUC) of edoxaban remained unaffected. However, its peak concentrations (Cmax) were decreased by 18.5%-28.1%. Similarly, co- administration of edoxaban with BBCT or HRHDT slightly decreased the AUC of M4 and reduced its C max by 16.8%-27.1%. Results revealed that BBCT and HRHDT had a minor impact on pharmacokinetics of edoxaban and M4. Despite alterations in systemic exposures, all pharmacodynamic parameters of edoxaban derived from activated partial thromboplastin time and prothrombin time were equivalent irrespective of herbal medicine co- administration. Conclusions: These findings contribute to our understanding of potential interactions between conventional anticoagulants and traditional herbal medicines, highlighting the need for comprehensive evaluation in clinical practice.

Retinal degenerative diseases, such as retinitis pigmentosa (RP) and age‑related macular degeneration (AMD), lead to progressive vision loss through photoreceptor degeneration; RP begins with the gradual loss of peripheral rods, whereas AMD causes central‑vision loss mainly because macular cones and parafoveal rods degenerate. The neural retina leucine zipper (NRL) directs rod photoreceptor differentiation, and its disruption has been linked to upregulated cone-specific markers in rods. This study investigates the therapeutic potential of a cell-penetrating asymmetric small interfering RNA targeting NRL (cp-asiNRL) to induce rod-to-cone conversion and mitigate retinal degeneration. cp-asiNRL was administered intravitreally to C57BL/6J wild-type (WT), neovascular AMD (nAMD), and RP (RhoP23H/+) mouse models. Subsequent analyses included cone marker expression levels and electroretinographic evaluations, and single-cell RNA sequencing. Administration of cp-asiNRL suppressed NRL expression, increased cone marker expression, and improved retinal function in both WT and nAMD models. In RP mice, cone marker expression was also elevated, although functional improvements were comparatively modest, likely reflecting the advanced disease stage. Single-cell RNA sequencing revealed a rod-to-cone-like transdifferentiation, suggesting that cp-asiNRL-mediated NRL knockdown partially preserved photoreceptor integrity. cp-asiNRL-mediated NRL silencing shows considerable promise as a therapeutic intervention for retinal degenerative conditions. By promoting rod-to-cone transdifferentiation and supporting photoreceptor survival, this approach may offer a novel strategy for vision preservation.

Introduction:Polycystic ovary syndrome (PCOS) is associated with an increased risk of non-alcoholic fatty liver disease (NAFLD). With the introduction of the new definition of metabolic dysfunction-associated fatty liver disease (MAFLD), there has been a lack of studies investigating the prevalence and clinical characteristics of PCOS and its phenotypes, including hyperandrogenism (HA), oligoanovulation (OA), and polycystic ovarian morphology (PCO) in association with MAFLD. The aim of this study is to explore MAFLD prevalence in young women with PCOS and determine the independent impact of PCOS phenotypes on MAFLD. Methods:This cross-sectional study included 1,422 women with PCOS diagnosed using the Rotterdam criteria, the presence of at least two of three diagnostic criteria: 1) hyperandrogenism (HA), 2) oligoanovulation (OA), and 3) polycystic ovary morphology (PCO). Results:Among women with PCOS, 31.2% had NAFLD, and 65.1% of them were diagnosed with MAFLD. In PCOS phenotypes, MAFLD prevalence was 25.1% for HA+OA+PCO, 27.6% for HA+OA, 8.8% for HA+PCO, and 13.0% for OA+PCO. Women with PCOS and HA+OA+PCO had higher odds of MAFLD (OR [95% CI] of 1.47 [1.04-2.09]), as did those with HA+OA (1.87 [1.18-2.96]), after adjusting for demographic and clinical factors. However, the association between women with PCOS and HA+PCO and MAFLD was not statistically significant (0.51 [0.21-1.24]). Discussion:In women with PCOS, both HA+OA+PCO and HA+OA phenotypes were independently associated with MAFLD. HA and OA may contribute independently to the higher prevalence of MAFLD in these individuals.

We study community detection in multiple networks whose nodes and edges are jointly correlated, a scenario that arises naturally in social platforms. Extending the classical Stochastic Block Model (SBM) and its contextual counterpart (CSBM), we introduce the correlated CSBMs, which incorporates both structural and attribute correlations. To build intuition, we first analyze correlated Gaussian Mixture Models, where a distance-based estimator can align nodes across graphs (without edges) even when community labels are initially unknown. For correlated CSBMs, we develop a two-step procedure that matches most nodes via $k$-core matching based on edge information, then refines the matching for remaining nodes by leveraging their attributes, enabling perfect alignment under suitable conditions. By aligning and combining graphs, we identify regimes where community detection is infeasible in a single graph but becomes possible when side information from correlated graphs is incorporated, illustrating how interplay between graph matching and community recovery enhances the inference on graphs.

Purpose: To evaluate the diagnostic performance of the American College of Radiology (ACR) Bone Reporting and Data System (Bone-RADS) in diagnosis of malignant tumors of the appendicular bone on conventional radiographs. Methods: Primary and secondary tumors of appendicular bone in patients who underwent radiographic and MRI examinations were classified into benign, intermediate, and malignant using a reference standard of histopathology, imaging follow-up, or clinical-radiologic consensus. Two radiologists assessed five radiographic features (margin, periosteal reaction, endosteal erosion, pathologic fracture, and extra-osseous mass), scored point total (points from radiographic features and a history of cancer), and assigned Bone-RADS categories. The diagnostic performance of Bone-RADS and interreader agreements were calculated. Results: A total of 778 patients (507 benign, 45 intermediate, and 226 malignant tumors) were included. BoneRADS showed high discrimination performance, with areas under the receiver-operating characteristics curve of 0.940-0.957 for point total and 0.895-0.900 for categorization. Bone-RADS had high sensitivity (95.2 %-99.1 %) and negative predictive value (NPV) (96.4 %-99.5 %), but relatively low specificity (65.0 %-68.6 %) and positive predictive value (PPV) (53.7 %-61.9 %). Interreader agreements were good to excellent for Bone-RADS point total (ICC = 0.850), categorization (k = 0.739), and most of the radiographic features (k = 0.621-0.822), except for endosteal erosion (k = 0.537) and extra-osseous mass (k = 0.234). Conclusion: In diagnosis of malignant bone tumors, ACR Bone-RADS showed high discrimination performance, with high sensitivity and NPV, but relatively low specificity and PPV. Nevertheless, relatively low interobserver agreement in some radiographic features and the consensus-based points system in Bone-RADS warrant further research and possible updates.

We consider the problem of detecting the presence of a signal in a rank-one spiked Wigner model. For general non-Gaussian noise, assuming that the signal is drawn from the Rademacher prior, we prove that the log likelihood ratio (LR) of the spiked model against the null model converges to a Gaussian when the signal-to-noise ratio is below a certain threshold. The threshold is optimal in the sense that the reliable detection is possible by a transformed principal component analysis (PCA) above it. From the mean and the variance of the limiting Gaussian for the log LR, we compute the limit of the sum of the Type-I error and the Type-II error of the likelihood ratio test. We also prove similar results for a rank-one spiked IID model where the noise is asymmetric but the signal is symmetric.

Purpose:Predicting long-term anatomical responses in neovascular age-related macular degeneration patients is critical for patient-specific management. This study validates a generative deep learning model to predict 12-month posttreatment optical coherence tomography (OCT) images and evaluates the impact of incorporating clinical data on predictive performance.Methods:A total of 533 eyes from 513 treatment-na & iuml;ve neovascular age-related macular degeneration patients were analyzed. A conditional generative adversarial network served as the baseline model, generating 12-month OCT images using pretreatment OCT, fluorescein angiography, and indocyanine green angiography. We then sequentially added OCT after three loading doses, baseline visual acuity, treatment regimen (pro re nata or treat-and-extend), drug type, and switching events. The generated and patient OCT images were compared for intraretinal fluid, subretinal fluid, pigment epithelial detachment, and subretinal hyperreflective material, both qualitatively and quantitatively.Results:The baseline model achieved acceptable accuracy for 4 macular fluid compartments (range 0.74-0.96). Incorporating OCT after loading doses and other clinical parameters improved accuracy (range 0.91-0.98). With all the clinical inputs, the model achieved 92% accuracy in distinguishing wet macular status from dry macular status. The segmented fluid compartments in the generated images correlated positively with those in the patient images.Conclusion:Integrating clinical and treatment data, particularly OCT data after loading doses, significantly enhanced the 12-month predictive performance of conditional generative adversarial networks. This approach can help clinicians anticipate anatomical outcomes and guide personalized, long-term neovascular age-related macular degeneration treatment strategies.

We investigate the mechanisms of self-distillation in multi-class classification, particularly in the context of linear probing with fixed feature extractors where traditional feature learning explanations do not apply. Our theoretical analysis reveals that multi-round self-distillation effectively performs label averaging among instances with high feature correlations, governed by the eigenvectors of the Gram matrix derived from input features. This process leads to clustered predictions and improved generalization, mitigating the impact of label noise by reducing the model's reliance on potentially corrupted labels. We establish conditions under which multi-round self-distillation achieves 100\% population accuracy despite label noise. Furthermore, we introduce a novel, efficient single-round self-distillation method using refined partial labels from the teacher's top two softmax outputs, referred to as the PLL student model. This approach replicates the benefits of multi-round distillation in a single round, achieving comparable or superior performance--especially in high-noise scenarios--while significantly reducing computational cost.

Test-Time Adaptation (TTA) adjusts models using unlabeled test data to handle dynamic distribution shifts. However, existing methods rely on frequent adaptation and high computational cost, making them unsuitable for resource-constrained edge environments. To address this, we propose SNAP, a sparse TTA framework that reduces adaptation frequency and data usage while preserving accuracy. SNAP maintains competitive accuracy even when adapting based on only 1\% of the incoming data stream, demonstrating its robustness under infrequent updates. Our method introduces two key components: (i) Class and Domain Representative Memory (CnDRM), which identifies and stores a small set of samples that are representative of both class and domain characteristics to support efficient adaptation with limited data; and (ii) Inference-only Batch-aware Memory Normalization (IoBMN), which dynamically adjusts normalization statistics at inference time by leveraging these representative samples, enabling efficient alignment to shifting target domains. Integrated with five state-of-the-art TTA algorithms, SNAP reduces latency by up to 93.12\%, while keeping the accuracy drop below 3.3\%, even across adaptation rates ranging from 1\% to 50\%. This demonstrates its strong potential for practical use on edge devices serving latency-sensitive applications. The source code is available at https://github.com/chahh9808/SNAP.

We study community detection in multiple networks with jointly correlated node attributes and edges. This setting arises naturally in applications such as social platforms, where a shared set of users may exhibit both correlated friendship patterns and correlated attributes across different platforms. Extending the classical Stochastic Block Model (SBM) and its contextual counterpart (Contextual SBM or CSBM), we introduce the correlated CSBM, which incorporates structural and attribute correlations across graphs. To build intuition, we first analyze correlated Gaussian Mixture Models, wherein only correlated node attributes are available without edges, and identify the conditions under which an estimator minimizing the distance between attributes achieves exact matching of nodes across the two databases. For the correlated CSBMs, we develop a two-step procedure that first applies k-core matching to most nodes using edge information, then refines the matching for the remaining unmatched nodes by leveraging their attributes with a distance-based estimator. We identify the conditions under which the algorithm recovers the exact node correspondence, enabling us to merge the correlated edges and average the correlated attributes for enhanced community detection. Crucially, by aligning and combining graphs, we identify regimes in which community detection is impossible in a single graph but becomes feasible when side information from correlated graphs is incorporated. Our results illustrate how the interplay between graph matching and community recovery can boost performance, broadening the scope of multi-graph, attribute-based community detection.

By investigating the correlation between the injection rate and pressure of subretinal tissue plasminogen activator (tPA) and air using a standard Viscous Fluid Control (VFC) system with a 38-gauge cannula, we aimed to establish guidelines for stable injections. We fabricated a retina mimicking model (RMM) with 0.25% agarose solution and an aluminum plate, and substituted submacular hemorrhage (SMH) and tPA with blood-mimicking fluid (BMF) and balanced salt solution (BSS), respectively. The diameter of the pre-bleb mimicking SMH in RMM was 1.30 +/- 0.16 cm, increasing to 1.98 +/- 0.24 cm and 1.83 +/- 0.22 cm after bleb propagation with BSS and air, respectively. BSS injection rates were 2.86 +/- 0.04 mu l/sec, 6.74 +/- 0.48 mu l/sec and 8.55 +/- 0.16 mu l/sec at 8, 12, and 16 psi, respectively. Air injection rates were 37.98 +/- 3.11 mu l/sec, 79.01 +/- 5.13 mu l/sec and 156.06 +/- 13.72 mu l/sec at 2, 3 and 4 psi, respectively. By experimenting with different pressures in the RMM, we found 12 psi to be the minimum for proper BSS injection and 2 psi for air. These findings provide crucial parameters for safer surgery to prevent irreversible damage.

Abstract Study question To compare clinical and surgical outcomes of robotic single-port myomectomy using da Vinci® SP and robotic single-site myomectomy with Xi or Si for fertility preservation. Summary answer The robotic single-port myomectomy (da Vinci® SP) might be feasible, even if the myoma is deep-seated, compared to robotic single-site myomectomy (Xi or Si system). What is known already Myomas are clinically apparent in about 25% of women and become symptomatic in their reproductive age. Myomectomy is the choice of treatment for women desiring uterine preservation, especially in deep-seated fibroid that can affect pregnancy. Although young women prefer minimal invasive surgical technique with less scar, single-port laparoscopic myomectomy has some limitations. Adoption of the da Vinci® system made it possible to overcome the weaknesses of laparoscopy. Owing to the ability of single-port laparoscopy to reduce pain and improve patient satisfaction, robotic single-site surgery was developed in order of the da Vinci® Si, Xi and the latest new SP system. Study design, size, duration We retrospectively reviewed medical records of 214 patients who underwent robotic single-port myomectomy (RSPM, n = 111) using the da Vinci® SP surgical system or robotic single-site myomectomy (RSSM, n = 103) with the da Vinci® Xi or Si system between October 2015 and September 2023 at Ewha Womans University Mokdong Hospital. Baseline characteristics and operative outcomes were compared and analyzed between the RSPM and RSSM groups. Participants/materials, setting, methods We assessed FIGO classification and maximum diameter of fibroid. The incision time, docking time, console time, wound closure time, total operation time, estimated blood loss (EBL), chopping time for specimen removal, the number and weight of removed myoma were measured. Main results and the role of chance Mean age (38.5±6.6 vs 37.3±6.6 years) and anthropometric index were not significantly different between RSPM and RSSM groups. Although the diameter (7.1±1.4 vs 7.2±2.0 cm), number (3.1±2.7 vs 2.5±2.2), and weight (149.1±107.1 vs 146.4±130.6 g) of myoma were not significantly different between the two groups, the more deep-seated myoma diagnosed by FIGO classification (2.9±1.5 vs 4.6±2.0, p<0.0001) could be resected in RSPM group. The incision (3.6±1.9 vs 7.3±4.4 min, p <0.0001) and docking time (2.7±1.4 vs 4.6±1.6 min. p<0.0001) were less in the group of RSPM compared to RSSM that was earlier implementation. Depending on the differences in da Vinci® SP and Xi or Si systems, the wound was slightly larger, requiring shorter chopping time (6.4±5.0 vs 8.4±7.2 min, p=0.022) and longer closure time (12.8±4.4 vs 10.9±3.6 min, p=0.001) in the RSPM group. Additionally, more complex surgeries with larger blood loss (264.1±140.2 vs 176.8±132.0 ml, p<0.0001), longer console time (60.5±33.2 vs 49.7±22.6 min, p=0.016) and hospitalization (4.6±0.9 vs 4.3±0.6 days, p=0.005) were performed in the RSPM group, but the difference in hemoglobin (2.5±1.0 vs 2.6±0.9 g/dL) before and after surgery was not significant. Limitations, reasons for caution All surgeries underwent using the da Vinci systems by one gynecologic surgeon who had experience performing almost 2,000 gynecologic robotic surgeries and the follow-up period was not enough to conclude the reproductive outcome. Wider implications of the findings The RSPM using the da Vinci® SP surgical system could be recommended for patient who planned future pregnancy. However, the optimal surgical technique should be selected for successful fertility preservation of reproductive surgery. Trial registration number 2023-03-038

Background and Objectives: Chemokines have various biological functions and potential roles in the development or progression of neuroinflammatory diseases. However, the specific pathogenic roles of chemokines in the major cause for vision loss among the elderly, the leading cause of blindness in older individuals, remain elusive. Chemokines interact with their receptors expressed in the endothelium and on leukocytes. The sulfation of tyrosine residues in chemokine receptors increases the strength of ligand–receptor interaction and modulates signaling. Therefore, in the present study, we aimed to construct a human recombinant sulfated CXCR3 peptide trap (hCXCR3-S2) and mouse recombinant sulfated CXCR3 peptide trap (mCXCR3-S2) to demonstrate in vivo effects in preventing choroidal neovascularization (CNV) and chemotaxis. Materials and Methods: We generated expression vectors for mCXCR3-S2 and hCXCR3-S2 with GST domains and their respective cDNA sequences. Following overexpression in E. coli BL21 (DE3), we purified the fusion proteins from cell lysates using affinity chromatography. First, the impact of hCXCR3-S2 was validated in vitro. Subsequently, the in vivo efficacy of mCXCR3-S2 was investigated using a laser-induced CNV mouse model, a mouse model of neovascular age-related macular degeneration (AMD). Results: hCXCR3-S2 inhibited the migration and invasion of two human cancer cell lines. Intravitreal injection of mCXCR3-S2 attenuated CNV and macrophage recruitment in neovascular lesions of mouse models. These in vitro and in vivo effects were significantly stronger with CXCR3-S2 than with wild-type CXCR3 peptides. Conclusion: These findings demonstrate that the sulfated form of the CXCR3 peptide trap is a valuable tool that could be supplemented with antivascular endothelial growth factors in AMD treatment.

BACKGROUND AND OBJECTIVE:This study aimed to assess and compare the pharmacokinetics, safety, and tolerability of a fixed-dose combination product (FDCP) comprising four different drugs (two antihypertensive drugs, amlodipine and losartan, and two lipid-lowering agents, ezetimibe and rosuvastatin) with their separate tablets. METHODS:A total of 60 participants were enrolled in this open-label, randomized, single-dose crossover study. Each participant received a single dose of FDCP and individual tablets during each period, with a 14-day washout period between the periods. The pharmacokinetic parameters of amlodipine, losartan, EXP3174 (an active metabolite of losartan), rosuvastatin, free ezetimibe, and total ezetimibe were evaluated and compared. RESULTS:The pharmacokinetic profiles of amlodipine, losartan, rosuvastatin, and ezetimibe after administration of the individual products were similar to those of FDCP. The geometric mean ratios and 90% confidence intervals for maximum concentration (Cmax) and area under the curve (AUC) of FDCP to individual tablets were within 0.8-1.25 for all six analytes. No clinically relevant changes were observed in the vital signs or physical, biochemical, hematological, electrocardiographic, or urinalysis findings during the study, and no serious adverse events were reported. CONCLUSION:This study demonstrated that a newly developed FDCP containing amlodipine, losartan, ezetimibe, and rosuvastatin exhibited pharmacokinetic equivalence with the individual products and met the regulatory criteria. Both formulations were well tolerated. CLINICAL TRIAL REGISTRATION:This trial (NCT04322266) was retrospectively registered on 9 September 2019.

Dataset distillation aims to synthesize a small number of images per class (IPC) from a large dataset to approximate full dataset training with minimal performance loss. While effective in very small IPC ranges, many distillation methods become less effective, even underperforming random sample selection, as IPC increases. Our examination of state-of-the-art trajectory-matching based distillation methods across various IPC scales reveals that these methods struggle to incorporate the complex, rare features of harder samples into the synthetic dataset even with the increased IPC, resulting in a persistent coverage gap between easy and hard test samples. Motivated by such observations, we introduce SelMatch, a novel distillation method that effectively scales with IPC. SelMatch uses selection-based initialization and partial updates through trajectory matching to manage the synthetic dataset's desired difficulty level tailored to IPC scales. When tested on CIFAR-10/100 and TinyImageNet, SelMatch consistently outperforms leading selection-only and distillation-only methods across subset ratios from 5% to 30%.

Background: In recent years, there has been considerable interest in the therapeutic targeting of tumor-associated macrophages (TAMs) to modulate the tumor microenvironment (TME), resulting in antitumoral phenotypes. However, key mediators suitable for TAM-mediated remodeling of the TME remain poorly understood. Methods: In this study, we used single-cell RNA sequencing analyses to analyze the landscape of the TME modulated by TAMs in terms of a protumor microenvironment during early tumor development. Results: Our data revealed that the depletion of TAMs leads to a decreased epithelial-to-mesenchymal transition (EMT) signature in cancer cells and a distinct transcriptional state characterized by CD8+ T cell activation. Moreover, notable alterations in gene expression were observed upon the depletion of TAMs, identifying Galectin-1 (Gal-1) as a crucial molecular factor responsible for the observed effect. Gal-1 inhibition reversed immune suppression via the reinvigoration of CD8+ T cells, impairing tumor growth and potentiating immune checkpoint inhibitors in breast tumor models. Conclusion: These results provide comprehensive insights into TAM-mediated early tumor microenvironments and reveal immune evasion mechanisms that can be targeted by Gal-1 to induce antitumor immune responses.