OBJECTIVE To construct a predictive model for evaluating the efficacy of a triple antiemetic regimen （neurokinin- 1 receptor antagonist+5-hydroxytryptamine 3 receptor antagonist+dexamethasone） for preventing nausea and vomiting induced by highly emetogenic chemotherapy （HEC） based on interpretable deep learning algorithms. METHODS Clinical data of cancer patients who received HEC and were treated with the standard triple antiemetic regimen in the oncology department of Tianjin Medical University General Hospital from January 2018 to December 2022 were collected retrospectively. Demographic， clinical and metabolism-related variables were integrated. After data pre-processing， two deep learning algorithms （deep random forest and dense neural network） and four machine learning algorithms （support vector machine， categorical boosting， random forest and decision tree） were used to build predictive models. Subsequently， model performance evaluation and model interpretability analysis were conducted. RESULTS Among the six candidate models， the deep random forest model demonstrated the best predictive performance on the test set， with an area under the receiver operating characteristic curve of 0.850， an accuracy of 0.911， a precision of 0.805， a recall of 0.783， an F1 score of 0.793， and a Brier score of 0.075. Interpretability analysis revealed that creatinine clearance rate （Ccr） was the key predictive factor， and low Ccr levels， female gender， younger age， highly emetogenic drugs （particularly cisplatin-containing chemotherapy regimens）， and anticipatory nausea and vomiting were positively correlated with the risk of HEC-related nausea and vomiting. CONCLUSIONS The deep random forest model exhibits the best performance in predicting the efficacy of triple antiemetic regimen for preventing HEC-related nausea and vomiting. The key predictors in this model primarily include Ccr，anticipatory nausea and vomiting， gender， age， and highly emetogenic drugs.

ObjectiveDiscriminating atypical hepatocellular carcinoma (HCC) from other malignancies in liver nodules classified as Liver Imaging Reporting and Data System category M (LR-M) remains a significant diagnostic challenge on conventional ultrasound examination. The LR-M category, originally intended to capture non-HCC malignancies, paradoxically contains up to 63% of atypical HCCs that deviate from classic enhancement patterns, leading to potential misdiagnosis and suboptimal treatment planning. While deep learning has shown promise in HCC diagnosis, most existing models rely exclusively on single-modality ultrasound, overlooking the diagnostic benefits of integrating complementary information from multiple imaging sources. To address this gap, we propose a novel attention-weighted tri-modal ultrasound network (TUS-Net) that integrates contrast-enhanced ultrasound (CEUS), B-mode ultrasound (BUS), and time-intensity curves (TICs) to improve diagnostic accuracy for these clinically challenging lesions. MethodsOur framework incorporates a three-dimensional convolutional neural network (C3D) backbone to extract spatiotemporal features from CEUS videos, capturing dynamic vascular patterns critical for lesion characterization. To effectively fuse complementary modalities, we introduce a dual-channel feature fusion module (DCFFM) that adaptively combines features from CEUS and BUS through channel-wise attention mechanisms, allowing the model to dynamically weigh the contribution of each modality based on diagnostic relevance. Additionally, we propose a temporal intensity feature fusion module (TIFFM) that leverages quantitative hemodynamic information from TICs to guide the model’s attention toward diagnostically critical temporal phases, such as arterial wash-in and portal venous washout. The model is further enhanced by automated lesion localization using YOLOX and class activation mapping for interpretability, ensuring that predictions align with clinically meaningful imaging features. ResultsEvaluated on a tri-modal ultrasound dataset comprising 161 patients with pathologically confirmed LR-M nodules (131 atypical HCC and 30 non-HCC malignancies), our model achieved an accuracy of 86.83%, a sensitivity of 92.50%, a specificity of 75.50%, and an AUC of 89.32% in screening atypical HCC. Compared to single-modality baselines, TUS-Net demonstrated superior specificity, a clinically critical metric given the higher risk associated with misclassifying non-HCC malignancies. Ablation studies confirmed the contribution of each module, with the full model outperforming both standard C3D and 3D ResNet backbones integrated with attention mechanisms. A reader study involving junior and senior radiologists further validated the clinical utility of AI assistance, showing consistent improvements in specificity and inter-reader consistency, particularly for less experienced clinicians. ConclusionThese results surpass existing benchmark models and demonstrate the potential of our approach to enhance diagnostic precision in clinically specific cases. By intelligently fusing multi-modal ultrasound data with attention-guided mechanisms, TUS-Net offers a reliable and interpretable tool that holds promise for improving the non-invasive diagnosis of atypical HCC in challenging LR-M liver nodules.

ObjectiveDiscriminating atypical hepatocellular carcinoma (HCC) from other malignancies in liver nodules classified as Liver Imaging Reporting and Data System category M (LR-M) remains a significant diagnostic challenge on conventional ultrasound examination. The LR-M category, originally intended to capture non-HCC malignancies, paradoxically contains up to 63% of atypical HCCs that deviate from classic enhancement patterns, leading to potential misdiagnosis and suboptimal treatment planning. While deep learning has shown promise in HCC diagnosis, most existing models rely exclusively on single-modality ultrasound, overlooking the diagnostic benefits of integrating complementary information from multiple imaging sources. To address this gap, we propose a novel attention-weighted tri-modal ultrasound network (TUS-Net) that integrates contrast-enhanced ultrasound (CEUS), B-mode ultrasound (BUS), and time-intensity curves (TICs) to improve diagnostic accuracy for these clinically challenging lesions. MethodsOur framework incorporates a three-dimensional convolutional neural network (C3D) backbone to extract spatiotemporal features from CEUS videos, capturing dynamic vascular patterns critical for lesion characterization. To effectively fuse complementary modalities, we introduce a dual-channel feature fusion module (DCFFM) that adaptively combines features from CEUS and BUS through channel-wise attention mechanisms, allowing the model to dynamically weigh the contribution of each modality based on diagnostic relevance. Additionally, we propose a temporal intensity feature fusion module (TIFFM) that leverages quantitative hemodynamic information from TICs to guide the model’s attention toward diagnostically critical temporal phases, such as arterial wash-in and portal venous washout. The model is further enhanced by automated lesion localization using YOLOX and class activation mapping for interpretability, ensuring that predictions align with clinically meaningful imaging features. ResultsEvaluated on a tri-modal ultrasound dataset comprising 161 patients with pathologically confirmed LR-M nodules (131 atypical HCC and 30 non-HCC malignancies), our model achieved an accuracy of 86.83%, a sensitivity of 92.50%, a specificity of 75.50%, and an AUC of 89.32% in screening atypical HCC. Compared to single-modality baselines, TUS-Net demonstrated superior specificity, a clinically critical metric given the higher risk associated with misclassifying non-HCC malignancies. Ablation studies confirmed the contribution of each module, with the full model outperforming both standard C3D and 3D ResNet backbones integrated with attention mechanisms. A reader study involving junior and senior radiologists further validated the clinical utility of AI assistance, showing consistent improvements in specificity and inter-reader consistency, particularly for less experienced clinicians. ConclusionThese results surpass existing benchmark models and demonstrate the potential of our approach to enhance diagnostic precision in clinically specific cases. By intelligently fusing multi-modal ultrasound data with attention-guided mechanisms, TUS-Net offers a reliable and interpretable tool that holds promise for improving the non-invasive diagnosis of atypical HCC in challenging LR-M liver nodules.

Diabetic foot ulcers (DFUs) are among the most serious and common complications of diabetes mellitus. They significantly affect patients’ quality of life and impose a substantial economic burden on healthcare systems worldwide. In Korea, the prevalence of diabetes and related complications, such as DFUs, has been increasing, reflecting a broader global trend. DFUs are associated with severe complications, including infections, neuropathy, and peripheral arterial disease, often leading to amputation. In Korea, diabetic foot complications are a major cause of non-traumatic lower-extremity amputations, with high mortality rates following amputation. DFUs also significantly reduce patients’ quality of life and increase healthcare costs. The management of DFUs requires a multidisciplinary approach that integrates medical, surgical, and advanced therapeutic interventions to prevent severe outcomes, such as amputation. This comprehensive review of DFU management in patients with diabetes was developed in collaboration with the Diabetic Study Group of the Korean Diabetes Association and Korean Society for Diabetic Foot. This review examines the epidemiology, clinical significance, diagnosis, and evidence-based treatment of DFUs.

This study aimed to determine the prevalence of human epidermal growth factor receptor 2 (HER2)–ultralow breast cancer among cases initially classified as HER2 immunohistochemistry (IHC) 0 and assess interobserver variability in interpreting low-level HER2 expression. Methods: In this multicenter retrospective study, all invasive breast cancer cases diagnosed between January and December 2022 across 10 Korean institutions were retrieved. Institutional pathologists reexamined HER2 IHC slides originally reported as IHC 0 according to the 2018 American Society of Clinical Oncology/College of American Pathologists guidelines and reclassified them as HER2-null (0), HER2-ultralow (0+), or HER2-low (1+). Slides from 10% of HER2-null and HER2-ultralow cases were digitized for central review and independently assessed by two pathologists, with discrepancies resolved by consensus. Results: Among 8,026 cases, 2,836 cases (35.5%) were initially reported as IHC 0. Upon re-review, 1,673 (59.0%), 1,139 (40.2%), and 24 (0.8%) cases were reclassified as HER2-null, HER2-ultralow, and HER2-low, respectively. The prevalence of HER2-ultralow breast cancer varied considerably across institutions (23.7%–78.1%). Central review of 268 digitized cases showed concordance in 193 cases (72.0%). Among the 75 discordant cases, 54 tumors (72.0%) were upgraded from HER2-null to HER2-ultralow, and 18 (24.0%) tumors were upgraded from HER2-ultralow to HER2-low. Furthermore, two tumors (2.7%) were downgraded from HER2-ultralow to HER2-null. Conclusions: Approximately 40% of cases initially categorized as IHC 0 were reclassified as HER2-ultralow. The substantial inter-institutional variability observed in interpreting low-level HER2 expression highlights the need for standardized training and quality assurance to ensure accurate identification of patients eligible for HER2-targeted antibody–drug conjugates.

Objective: To evaluate the accuracy and reasoning capabilities of large multimodal language models compared with those of neuroradiology subspecialty-trained radiologists in neuroradiology case interpretation. Materials and Methods: This experimental study used custom-made 401 radiologic quizzes derived from articles published in RadioGraphics covering neuroradiology and head and neck topics (October 2020 to February 2024). We prompted the GPT-4 Turbo with Vision (GPT-4V), GPT-4 Omni, Gemini Flash, and Claude models to provide the top three differential diagnoses with a rationale and describe examination characteristics such as imaging modality, sequence, use of contrast, image plane, and body part. The temperature was adjusted to 0 and 1 (T1). Two neuroradiologists answered the same questions.The accuracies of the large language models (LLMs) and the neuroradiologists were compared using generalized estimating equations. Three neuroradiologists assessed the rationale provided by the LLMs for their differential diagnoses using four-point scales, separately for specific lesion locations and imaging findings, and evaluated the presence of hallucinations and the overall acceptability of the responses. Results: Top-3 accuracy (i.e., correct answers present among top-3 differential diagnoses) of LLMs ranged from 29.9% (120 of 401) to 49.4% (198 of 401, obtained with GPT-4V in the T1 setting), while radiologists achieved 80.3% (322 of 401) and 68.3% (274 of 401), respectively (P < 0.001). Regarding the rationale for differential diagnoses, GPT-4V (T1) accurately identified both the specific lesion location and imaging findings in 30.7% (123 of 401) and 12.9% (16 of 124) of cases without textual clinical history. Hallucinations occurred in 4.5% (18 of 401), and only 29.4% (118 of 401) of the LLM-generated analyses were deemed acceptable. GPT-4V (T1) demonstrated high accuracy in identifying the imaging modality (97.4% [800 of 821]) and scanned body parts (92.2% [756 of 820]). Conclusion LLMs remarkably underperformed compared with neuroradiologists and showed unsatisfactory reasoning for their differential diagnoses, with performance declining further in cases without textual input of clinical history. These findings highlight the limitations of current multimodal LLMs in neuroradiological interpretation and their reliance on text input.

In patients with differentiated thyroid cancer (DTC), initial recurrence risk stratification based on clinical, histopathological, and perioperative data remains the key determinant for guiding management strategies during the first 1-2 years post-treatment. However, the adoption of ongoing risk stratification (ORS), which dynamically reassesses risk by integrating longitudinal clinical data and treatment response, enables more precise long-term prognostic assessment and facilitates highly individualized management. Building upon recent guidelines, the 2026 KTA guideline has been further refined by incorporating robust evidence from large-scale national cohorts and comprehensive systematic reviews. These updated recommendations outline contemporary concepts of ORS, risk-adapted TSH suppression targets, optimized surveillance modalities for recurrence detection, and disease-specific long-term follow-up strategies. Reflecting the paradigm shift toward de-escalated treatment, this revision integrates evolved perspectives on TSH suppression intensity, the clinical interpretation of thyroglobulin levels, and tailored follow-up intervals. These evidence-based recommendations aim to minimize unnecessary treatment and excessive surveillance in the large proportion of patients with excellent prognosis after initial therapy, while ensuring that each patient receives appropriately tailored and effective long-term management.

Purpose: The aim of this study was to investigate the feasibility of an intelligent handheld ultrasound (US) device for assisting non-expert general practitioners (GPs) in detecting carotid plaques (CPs) in community populations. Methods: This prospective parallel controlled trial recruited 111 consecutive community residents. All of them underwent examinations by non-expert GPs and specialist doctors using handheld US devices (setting A, setting B, and setting C). The results of setting C with specialist doctors were considered the gold standard. Carotid intima-media thickness (CIMT) and the features of CPs were measured and recorded. The diagnostic performance of GPs in distinguishing CPs was evaluated using a receiver operating characteristic curve. Inter-observer agreement was compared using the intragroup correlation coefficient (ICC). Questionnaires were completed to evaluate clinical benefits. Results: Among the 111 community residents, 80, 96, and 112 CPs were detected in settings A, B, and C, respectively. Setting B exhibited better diagnostic performance than setting A for detecting CPs (area under the curve, 0.856 vs. 0.749; P<0.01). Setting B had better consistency with setting C than setting A in CIMT measurement and the assessment of CPs (ICC, 0.731 to 0.923). Moreover, measurements in setting B required less time than the other two settings (44.59 seconds vs. 108.87 seconds vs. 126.13 seconds, both P<0.01). Conclusion Using an intelligent handheld US device, GPs can perform CP screening and achieve a diagnostic capability comparable to that of specialist doctors.