European Radiology

Background: Comprehensive evaluations of lung nodule detection methods remain limited, particularly regarding data from clinical routine, novel photon-counting CT (PCCT) technology, and the comparison of academic and commercial models. Materials and Methods: This retrospective study collected 1860 routine PCCT scans acquired at Hannover Medical School from 2021 to 2024. We compared four deep-learning-based lung nodule detection models, which were developed for conventional CT, namely the publicly available TotalSegmentator, nnDetection, and RadYOLO methods, as well as a commercial lung computer-aided detection (CAD) system. We evaluated the inter-model agreement in an unannotated cohort (n=1699), sensitivity in a smaller cohort of 25 manually annotated scans, and false positives as detections in 136 report-negative scans. Results: In the unannotated cohort, the commercial CAD system detected 7071, TotalSegmentator 4264, nnDetection 5254, and RadYOLO 3122 nodules. Overall, the inter-model agreement was low, and the models produced substantially different candidate nodules. Approximately 1300 nodules were detected by all models, representing 18.7% - 42.8% of each model's detections. 15.4% - 37.4% of a model's detections did not match any other model. The models showed moderate sensitivity (between 0.71, 95%-CI [0.64, 0.77], and 0.83, 95%-CI [0.77, 0.88]), and TotalSegmentator and RadYOLO produced few false positives per scan (both 0.85, 95%-CI [0.70, 1.02]). Conclusion: The lung nodules detected by four deep learning models showed high variation and a lack of consensus on clinical PCCT data. These findings highlight the importance of model benchmarking and local validation when deploying models in real-world clinical workflows.

Computed tomography (CT) interpretation requires recognition of the anatomical level of each slice. Classification may facilitate preprocessing and AI-assisted fracture diagnosis. We developed deep learning models to classify craniofacial CT slices into six anatomical categories. A total of 11,769 images were collected from a multicenter database involving 10 Japanese university hospitals and the public Head-Neck-Radiomics-HN1 dataset. Patients were divided into training (70%) and evaluation (30%) sets. Five CNN architectures were trained from scratch, and hard- and soft-voting ensembles were evaluated. Individual-model accuracy ranged from 94.72% to 96.09%; the hard-voting ensemble achieved 96.46% accuracy (Cohen's k = 0.954). Although performance differed among individual models (Cochran's Q = 99.29, p < 0.001), the ensemble did not significantly outperform EfficientNetV2B1 (McNemar's p = 0.163). Errors occurred mainly between adjacent anatomical regions. Deep learning can classify craniofacial CT slices accurately and may support automated dataset organization and AI-assisted fracture diagnosis. External validation is needed.

Objectives: To determine how strongly clinical referral text, rather than image content, drives the decisions of multimodal large language models (LLMs) differentiating glioma from brain abscess on MRI. Materials and methods: This retrospective single-center study included 60 contrast-enhanced T1-weighted brain MRI examinations (30 gliomas, 30 abscesses; reference-standard diagnosis). One axial slice per case was paired with five referral texts (explicitly inflammatory to explicitly tumor-suggestive) under three prompt instructions and submitted 20 times each to GPT-4o, Claude Sonnet 5, Gemini 3.5 Flash, and Qwen 3.6 (72,000 answers). The primary outcome was the text-driven span (difference in tumor-call frequency between explicit inflammatory and explicit tumor text). Two blinded radiologists classified the images without text; sensitivity, specificity, and accuracy were compared descriptively. Results: Patients (29 women) had a mean age of 55.8 +/- 16.4 years. With neutral text, models answered tumor in 45-67% of queries (specificity, 0.38-0.78), versus 52% and 60% for radiologists (specificity, 0.77 and 0.63). The text-driven span was 0.79-1.00 (p < 0.001 for all models), whereas confirmed gliomas and abscesses differed by only 0.01-0.10 in three models and 0.25 in Gemini 3.5 Flash. Subtle wording produced at least 83% of the span in three models. An instruction to ignore the text reduced the span to 0.48-0.70 in three models and to 0.04 in Gemini 3.5 Flash, partly through reversion to a tumor default. Model accuracy was 0.52-0.73 versus 0.78 and 0.73 for radiologists. Conclusion: Multimodal LLMs deferred to referral text over imaging evidence in differentiating glioma from brain abscess, and prompt instructions did not reliably correct this.

Faster ultrahigh-resolution (UHR) photon-counting CT acquisitions facilitate breath-hold imaging, but identical reconstruction settings may conceal differences in spatial resolution. To assess spatial resolution and depiction of small airways and venous stents across helical UHR acquisition configurations. Stationary wire, airway, and venous-stent phantoms underwent one acquisition per configuration using five configurations on a dual-source scanner at 120 kVp. Reconstructions used 0.2- and 0.4-mm thicknesses, the same sharp kernel, and iterative reconstruction strength 1. We measured the full-width at half-maximum of the slice sensitivity profile (SSP) and spatial frequency at 20% of the modulation transfer function (MTF20). Rank-based F tests with Holm adjustment assessed configuration differences in resolution and variability across slices. Airway detection used a profile contrast-to-noise criterion; stent assessment was qualitative. Calculated acquisition time for 400-mm coverage decreased from 9.80 to 1.30 seconds between the slowest and fastest configurations. At 0.2-mm reconstruction thickness, median SSP increased from 0.316 to 0.496 mm, and median MTF20 decreased from 2.870 to 2.059 lp/mm (both P<.001). Interquartile ranges increased from 0.047 to 0.197 mm for SSP and from 0.256 to 0.654 lp/mm for MTF20 (both variability P<.001). Increasing reconstruction thickness to 0.4 mm increased SSP by a median 0.221 mm across 2459 matched slice-wire measurements (P<.001). At 0.2-mm reconstruction thickness, 0.4- and 0.6-mm lumens with 0.2-mm walls were detected only in the slowest configuration. Faster configurations showed less distinct stent struts and greater adjacent-wall distortion. Faster helical UHR photon-counting CT configurations had poorer spatial resolution despite identical nominal slice thickness. Consistent acquisition and reconstruction settings are important for longitudinal comparisons.

Objectives To evaluate discrepancies between radiologists and referring physicians in the interpretation of diagnostic probabilities conveyed by common Japanese expressions used in radiology reports. Materials and Methods Fifteen expressions commonly used to convey diagnostic probabilities in radiology reports were selected by five attending radiologists. An anonymous online survey was distributed to 26 radiologists and 33 referring physicians at a single institution. Participants assigned a diagnostic probability (from 0% to 100% in 10% increments) to each expression. Differences in median probabilities between radiologists and referring physicians were analyzed using the Mann-Whitney U test. Variability in interpretation was assessed using interquartile range (IQR), with IQR [&ge;] 30% defined as substantial variability. Results All radiologists (26/26) and 97% of referring physicians (32/33) completed the survey. Seven of the 15 expressions showed statistically significant differences in interpretation between radiologists and referring physicians: "c/w" (80% vs. 60%; p = 0.031), the Japanese equivalents of "present" (100% vs. 100%; p = 0.024), "suspect of" (80% vs. 70%; p = 0.007), "possibly" (50% vs. 70%; p < 0.001), "unlikely" (10% vs. 10%; p = 0.022), "cannot be ruled out" (25% vs. 40%; p < 0.001), and "not inconsistent with" (80% vs. 90%; p = 0.008). Substantial variability in interpretation (IQR [&ge;] 30%) was observed for four expressions: "cannot be ruled out" (IQR, 10%-40%), "s/o" (IQR, 60%-90%), "c/w" (IQR, 50%-90%), and "not inconsistent with" (IQR, 70%-100%). Conclusion This study revealed that nearly half of the Japanese expressions commonly used to convey diagnostic probabilities were interpreted differently by radiologists and referring physicians. These discrepancies highlight the need for a standardized lexicon to improve clarity and consistency in radiology reporting.

Purpose: To quantify the sampling error of conventional 2D regions of interest (ROIs) compared with whole-organ 3D volumetric CT attenuation for opportunistic hepatic steatosis screening, and to evaluate its dependency on intraparenchymal spatial heterogeneity. Materials and Methods: In this retrospective computational study (CLAIM-compliant), 971 abdominal CT scans were analyzed across a multicenter derivation cohort (TotalSegmentator v2, n = 841) and an external validation cohort (Medical Segmentation Decathlon Task03, n = 130). Automated 3D liver segmentation was performed using nnU-Net, followed by an 8-mm spherical 3D erosion to eliminate peripheral partial volume effects and major hilar vessels. Three clinical sampling paradigms (1, 3, and 5 axial 2D ROIs of 15 mm diameter) were modeled through 500 stochastic Monte Carlo iterations per scan (485,500 total simulations). Absolute sampling error (|Delta HU| = |HU_ROI - HU_3D|) and diagnostic misclassification rates at the standard <= 40 HU steatosis threshold were computed. Results: In the derivation cohort, the conventional 3-ROI approach yielded a mean absolute error of 6.61 +/- 7.83 HU (median: 2.92 HU), significantly underperforming compared to the 5-ROI strategy (5.21 +/- 6.00 HU; Wilcoxon paired test, p < 0.001) but improving on the single-ROI approach (11.32 +/- 14.16 HU; p < 0.001). Sampling error correlated strongly with parenchymal spatial heterogeneity (Pearson r = 0.850, 95% CI: [0.830-0.868], p < 0.001). In borderline attenuation zones (35-45 HU), diagnostic misclassification reached 45.2% in heterogeneous livers. Findings were corroborated on external validation, confirming the reproducible decay of sampling error (|Delta HU| = 7.38 HU for 1 ROI vs. 5.14 HU for 3 ROI, p < 0.001). Conclusion: Manual 2D ROI sampling exhibits an intrinsic metrological uncertainty exceeding clinically relevant thresholds (5 HU) in heterogeneous livers, misclassifying nearly half of borderline cases. Opportunistic screening requires fully automated 3D volumetric extraction embedded into clinical PACS workflows.

Automated dental implant planning using artificial intelligence (AI) faces challenges in interpretability, safety assurance, and applicability beyond single-tooth cases. This paper presents a unified explainable framework for cone-beam computed tomography (CBCT) that handles both single-tooth and consecutive multiple missing teeth. The method combines nnU-Net-based anatomical segmentation with a transparent geometric optimization pipeline. Implant placement is formulated as a lexicographic multi-objective constrained optimization problem that enforces clinical safety margins while balancing bone engagement, mesiodistal spacing, and buccolingual position. For multi-tooth cases, a global coordination strategy ensures inter-implant parallelism, minimum spacing, and a common insertion path. The framework was validated on 20 single-tooth and 9 consecutive multi-tooth cases. The AI plans achieved a centroid distance of 1.60+/-1.05 mm for single-tooth cases and 1.22+/-0.32 mm for multi-tooth cases, with approximately 7 degrees axial angle difference. All plans achieved 100% compliance with predefined safety thresholds. In a Turing-style test, experts were unable to distinguish AI from human intelligence (HI) plans under blinded conditions. By decoupling perception from reasoning and embedding clinical knowledge as explicit constraints, the framework offers an interpretable and clinically trustworthy solution for computer-aided implant dentistry.

Deep-silicon photon-counting CT uses a detector architecture distinct from cadmium-telluride systems and has not been evaluated for pediatric head imaging. To compare attenuation-dependent performance of deep-silicon photon-counting CT (dSi-PCCT) with energy-integrating detector CT (EID-CT) and assess spectral and ultra-high-resolution (UHR) reconstructions from one acquisition. Graded-thickness cranial attenuator and pediatric head phantoms were scanned at a matched target volume CT dose index of 18 mGy. dSi-PCCT used 120 kVp and generated 50-, 65-, and 100-keV images; EID-CT used 80 and 100 kVp. We measured CT-number bias, noise, contrast, and contrast-to-noise ratio (CNR) across 0--9 mm of nominal skull thickness. Mixed-effects models assessed imaging-condition effects. We compared matched standard-resolution and UHR temporal-bone images and generated calcium- and water-density maps from the same dSi-PCCT scan. Noise increased by 4.54%, 3.07%, and 1.35% per millimeter at 50, 65, and 100 keV, respectively, versus 8.38% and 7.84% at 80 and 100 kVp. Corresponding CNR decreases were 4.17%, 3.25%, and 1.75% versus 8.94% and 8.54% (interaction {rho} < 0.001). Changes in absolute CT-number bias and contrast did not differ among conditions ({rho} = 0.92 and {rho} = 0.59). In the pediatric head phantom, model-estimated 50-keV CNR ratios were 1.299 (95% CI, 1.163--1.452) versus 80 kVp and 1.424 (95% CI, 1.275--1.592) versus 100 kVp. Across three temporal-bone edges, UHR transition widths were 58.6%-61.9% narrower than standard-resolution widths. dSi-PCCT maintained CNR more consistently as nominal skull thickness increased and provided complementary spectral and UHR pediatric head images from one acquisition.

Purpose: The aims of this study were to develop an externally validated deep learning vascular segmentation model for digital enhancement of vertebral artery dissection detection on CTA and to assess clinical utility of the model through a paired crossover reader study. Materials and Methods: This retrospective, IRB-approved study conducted from September 2024 to July 2026 included an internal training cohort of 84 manually segmented CTAs plus 17 CTAs from the RSNA Intracranial Aneurysm Challenge (101 CTAs total) and an external testing cohort of 40 CTAs (22 positive, 18 negative for vertebral artery dissection). A nnU-Net (version 2) model was trained using five-fold cross-validation with the entire training cohort. Technical performance was evaluated using the Dice similarity coefficient (DSC). A two-part crossover reader study included 10 readers who assessed diagnostic accuracy, confidence, and interpretation time with and without segmentation-based augmentation. Statistical analyses included McNemar's exact test (accuracy) and Wilcoxon signed-rank test (confidence, time), with P < .05 considered significant. Results: The model achieved a DSC of 0.96 (SD 0.01). In the reader study, diagnostic accuracy was lower with augmentation than without (0.72 vs 0.84, P < 0.001), while confidence (3.98 vs 3.96, P = .77) and interpretation time (137.3 vs 166.7 seconds, P = .54) did not differ significantly. Subjectively, 7 of 10 readers reported they would use the tool routinely in acute or trauma settings. Conclusion: Although the deep learning model demonstrated accurate segmentation of vascular structures, further validation is needed to establish its utility for enabling accurate diagnosis of VAD in fast-paced clinical settings.

To compare virtual noncontrast (VNC) error between deep-silicon photon-counting CT (PCCT) and dual-energy energy-integrating detector CT (EID-CT) in a matched phantom and characterize tissue-specific error in humans. This prospective exploratory study included a matched phantom experiment and pilot human evaluation. Prototype projection-domain water-density maps served as the VNC surrogate. Twenty-four vials containing 0.1--20 mg I/mL in deionized water or simulated blood were scanned with PCCT and EID-CT. Three men underwent true noncontrast and contrast-enhanced PCCT; 80 matched measurements were analyzed. Paired t tests and mixed-effects models assessed system differences and iodine associations. Three participants (mean age, 67 {+/-} 10 years; all men) were included. Mean VNC error was 29.1 HU lower with PCCT than EID-CT (95% CI, 15.5--42.7 HU; {rho} < 0.001). Error was not associated with iodine concentration for PCCT (slope, 0.03 HU per mg I/mL; 95% CI, -0.92 to 0.98; {rho} = 0.95) but increased for EID-CT (slope, 3.51 HU per mg I/mL; 95% CI, 2.56--4.46 HU; {rho} < 0.001). The slopes differed by 3.48 HU per mg I/mL ({rho} < 0.001). In humans, errors in five non-blood-pool tissues were within 5 HU. Error was not associated with iodine concentration among 61 non-blood-pool measurements (adjusted slope, -0.17 HU per mg I/mL; 95% CI, -1.22 to 0.88; {rho} = 0.75). PCCT provided VNC measurements with errors closer to zero and less dependent on iodine than EID-CT in the phantom. In humans, error was generally small in non-blood-pool tissues and was not associated with iodine concentration.

Background: Conventional performance metrics do not indicate whether artificial intelligence (AI)-platforms and radiologists are right or wrong on the same examinations, nor whether their non-agreement is more often beneficial or harmful. Objective: To develop, illustrate, and apply directional metrics quantifying non-agreement between radiologists and AI-platforms against a common reference standard, and to adapt a fragility-based framework for quantifying the statistical robustness of these metrics. Methods: We defined Positive Non-Agreement (PNA, AI-platform correct, radiologist incorrect), Negative Non-Agreement (NNA, radiologist correct, AI-platform incorrect), and Net Non-Agreement (NENA, their difference). Metrics were illustrated in three hypothetical scenarios of 1,000 examinations each and applied to three previously published AI-radiologist studies on pulmonary embolism and mammography. Bootstrap confidence intervals were calculated, and a reclassification fragility index (FI) quantified the minimum case reclassification needed to remove statistical support for NENA's direction. Results: In illustration, NENA was positive in two scenarios and negative in one. In the three published studies, NENA ranged from -4.18% to +10.74% and decreased consistently with increasing radiologist experience (junior residents, +6.27%; consultants, +1.45%), a gradient driven primarily by declining PNA with modest increase in NNA. The reclassification FI was defined for all three illustrative scenarios (range, 0.10%-1.70%) and for six of the seven comparisons from these studies (range, 0.34%-4.31%), but not for the weakest comparison, whose confidence interval for NENA included zero. Conclusions: PNA, NNA, and NENA provide a directional, case-level complement to conventional performance metrics for AI-radiologist non-agreement. The accompanying reclassification FI offers a statistically grounded, generalizable method for quantifying their robustness.

To evaluate whether eight-bin deep-silicon photon-counting CT can quantify iodine and gadolinium separately and distinguish radiopaque embolic beads from vascular contrast in tumor phantoms. In this phantom study, two-input decomposition of scanner-generated iodine and water images was compared with three- and five-material decomposition of five reconstructed energy-bin images in 56 common test vials, including nine simulated-blood preparations. One tantalum and one bismuth vial were also tested with five-material decomposition. A separately calibrated two-input method using scanner-generated 70-keV and iodine images was applied to four static 3D-printed tumor vascular models: iodine-only and gadolinium-only controls and two models with iodine-containing beads and gadolinium-filled channels. Concentration errors were summarized by bias and root mean square error (RMSE). For iodine, RMSE was 0.40, 0.73, and 1.12 mg I/mL with two-input, three-material, and five-material decomposition, respectively; corresponding gadolinium RMSEs were 0.57, 0.55, and 0.46 mg Gd/mL. Two-input biases were +0.01 mg I/mL and +0.29 mg Gd/mL. Five-material decomposition distinguished tantalum and bismuth in the separate test vials. Beads and vascular gadolinium were spatially separated in the tumor models. In the two bead-containing models, gadolinium measured 32.56 {+/-} 8.02 and 37.02 {+/-} 7.93 mg Gd/mL (mean {+/-} spatial SD), compared with the prepared 35.00 mg Gd/mL. Deep-silicon photon-counting CT separately visualized radiopaque embolic beads and quantified vascular gadolinium from a single acquisition. Recalibrated scanner-generated images supported quantitative dual-contrast imaging, while research energy-bin decomposition distinguished additional materials.

Introduction: Resection margins in oral squamous cell carcinoma (OSCC) are frequently inadequate. This study investigated the accuracy and diagnostic performance of ex-vivo 7T MRI against histology. Methods: Fifty OSCC specimens underwent ex-vivo 7T MRI. Tumours were segmented on corresponding histology and twice on MRI: blinded to histology (radiologist segmentation; clinical practice) and using histology (histology-MR segmentation; MR-based ground truth). Resection margins were measured in five regions and compared with histology. Results: Anterior and posterior margins were generally smaller on MRI than on histology (maximum mean difference -2.5 mm), whereas cranial, caudal and deep margins were generally larger on MRI (mean differences <1 mm). For close margins (<5 mm), sensitivity was 95% for histology-MR segmentations and 91% for radiologist segmentations, with 75% specificity for both. Conclusion: Ex-vivo 7T MRI accurately measures resection margins and detects close margins with high sensitivity. A future clinical trial will determine whether intraoperative implementation reduces inadequate margins and adjuvant treatment.

Background Large-scale retrospective studies often require researchers to adjudicate outcomes or cohort eligibility from unstructured clinical records. Manual review can provide reliable labels but is difficult to scale. We developed and evaluated a consensus workflow using multiple large language models (LLMs) to automatically assign high-confidence outcome labels while deferring ambiguous cases for manual review. Methods We included 6,718 brain MRI reports from 5,856 patients. The expert-adjudicated reference set included 680 reports from 675 patients: 372 normal (age-appropriate without significant intracranial pathology), 90 white matter hyperintensity-only (WMH-only; isolated nonspecific WMHs potentially ambiguous for binary classification), and 218 abnormal (other intracranial pathology) reports. We evaluated proprietary-inclusive (Qwen3/Kimi-Linear/GPT-5) and open-source (Qwen3/Kimi-Linear/DeepSeek-R1) ensembles against individual LLMs. Ensembles assigned a normal or abnormal label only when all models agreed; all other reports were classified as uncertain and deferred for manual review. Individual LLMs assigned normal or abnormal predictions and classified WMH-only predictions as uncertain. In the full cohort, we compared the prevalence of active central nervous system (CNS)-related International Classification of Diseases, Tenth Revision (ICD-10) diagnoses across the normal, uncertain, and abnormal groups generated by each ensemble. Results In the expert-adjudicated reference set, both ensembles achieved 100% precision for automatically assigned normal and abnormal reports. GPT-5 was the strongest individual LLM but had lower precision than either ensemble, with normal and abnormal precision of 98.9% and 94.1%, respectively, and 17 expert-label mismatches versus none for the consensus ensembles. In the full cohort, 1.6% of reports classified as normal were associated with CNS-related ICD-10 diagnoses, compared with 29.7% of uncertain and 51.3% of abnormal reports for the proprietary-inclusive ensemble; corresponding rates were 1.6%, 31.1%, and 51.6% for the open-source ensemble. Conclusions Requiring unanimous agreement across multiple LLMs enabled high-precision automatic adjudication while directing ambiguous cases to manual review. This strategy may reduce the burden of manual adjudication and support scalable construction of high-purity research cohorts from unstructured clinical text.

Background: Many radiology decisions, such as whether a report describes a finding or how urgently a study should be read, are choices among a fixed set of answers. Decision models return a probability for each permitted answer in one step, without generating text, and an open-weight model can run within a hospital so that patient data stay local. No decision model has yet been specialized for radiology and compared with large language models (LLMs). Methods: We fine-tuned the general-purpose decision models Kev-27B (Qwen3.8-27B backbone) and Kev-9B on 36,109 records from public radiology datasets to obtain RadKev-27B and RadKev-9B, and compared them with the Kev models, the hosted decision models Jev and OpenAI Decisions, and the LLM Qwen3.8-27B, on 14,142 held-out questions in fifteen tasks (report reading, case diagnosis and classification, radiological knowledge). The prespecified primary outcome was task-averaged accuracy relative to Kev-27B. Results: RadKev-27B exceeded Kev-27B by 4.9 percentage points (95% CI 3.8 to 6.2), with the largest gain in report error detection (22.6 points); RadKev-9B exceeded Kev-9B by 6.9 points (5.5 to 8.4) and did not differ detectably from Kev-27B, a model three times its size. RadKev-27B was more accurate than the hosted decision models Jev (3.0 points, 1.7 to 4.4) and OpenAI Decisions (4.8, 3.2 to 6.4), than Qwen3.8-27B without reasoning (4.7, 3.4 to 6.0) and with reasoning (2.8 points, 1.0 to 4.6) on 1,675 questions, at 132-fold lower latency (128 ms vs 16.9 s per question). At an error rate of 5%, RadKev-27B could answer 90.5% of questions and Kev-27B 76.4%. Conclusions: RadKev-27B was more accurate than two hosted decision models and an LLM of the same size, with or without reasoning, at a fraction of the latency of reasoning, and can be deployed within an institution for categorical radiology decisions after validation on local data.

Hard-tissue surgery requires precise spatial control, while conventional computer-assisted navigation presents guidance separately from the operative field. Augmented reality can bring navigation information closer to the surgeon, but how best to translate this information into intuitive intraoperative guidance remains underexplored. We therefore developed an augmented reality navigation system (ANS) that translates real-time deviations from the planned trajectory into co-located holographic action cues. Evaluation comprised typodont and ex vivo porcine experiments, a randomized crossover study in 40 dental trainees, and eight clinical cases. In porcine mandibles, ANS showed lower angular deviation than static and dynamic navigation (1.78 vs 6.42 and 4.02 degrees). Among trainees, accuracy was comparable with dynamic navigation, while workload was lower (29.18 vs 33.15), usability higher (73.13 vs 68.88), and visual-focus ratings improved (3.55 vs 2.28). Clinical cases supported feasibility. These findings support co-located action guidance as an alternative interface paradigm for surgical navigation.

Objectives: Intraoperative ultrasound (IOUS) is an increasingly valuable real-time imaging modality in neurosurgery, enabling visualization of brain parenchyma, tumor margins, ventricular anatomy, and vascular structures during cranial procedures. Despite its clinical relevance, the development of artificial intelligence (AI)-assisted interpretation tools for neurosurgical IOUS has been limited by the scarcity of large, annotated datasets. Manual pixel-level annotation of ultrasound video frames is labor-intensive, requires specialized neuroanatomical expertise, and scales poorly. To address this bottleneck, we developed a proof-of-concept interactive semi-automated annotation pipeline designed to explore the feasibility of scalable, expert-guided labeling for intraoperative neurosurgical IOUS. Methods: We implemented the system as a napari-based graphical interface coupled with a scikit-learn Random Forest (RF) pixel classifier. Real intraoperative neurosurgical ultrasound videos from five craniotomy cases were used as training material. A neurosurgical expert annotated a minimal set of representative frames using brush-based labeling to define anatomical classes of interest. For each labeled pixel, a feature vector was computed comprising raw intensity, multi-scale Gaussian filtering ({sigma} = 1, 2, 3), Laplacian of Gaussian, and Sobel edge magnitude. The RF classifier was trained on accumulated annotations using class-balanced weighting and out-of-bag (OOB) validation, then applied across full video sequences. An incremental accumulator module enabled multi-case training. Segmentation outputs were exported as TIFF overlays for expert quality review. Results: Across five intraoperative cases, sparse expert annotation of a small number of representative frames per case was sufficient to train case-specific Random Forest classifiers that propagated labels across the full video sequences. Out-of-bag estimation served as an internal coherence check during training, and the classifier additionally produced per-class probability maps that exposed regions of uncertainty at acoustic boundaries. On expert qualitative review, the auto-segmented frames were anatomically coherent, with the classifier reliably distinguishing anechoic fluid compartments from echogenic tissue interfaces; residual misclassification was concentrated in low-contrast boundary regions. Conclusions: This proof-of-concept study demonstrates that semi-automated, expert-guided annotation of intraoperative neurosurgical ultrasound is technically feasible using classical machine learning. By combining neurosurgical domain expertise with an interactive, incremental annotation workflow applied to real intraoperative acquisitions, this framework represents a foundational step toward scalable generation of labeled datasets for neurosurgical IOUS. Future directions include formal multi-rater validation, expanded multi-class anatomical labeling, and eventual migration to deep learning architectures as annotated datasets grow.

Background Accurate prediction of 90-day outcomes following ischemic stroke is crucial for optimizing clinical strategies. This study aimed to systematically review and meta-analyze studies using deep learning (DL) and hand-crafted radiomics (HCR) models for predicting 90-day modified Rankin Scale (mRS) outcomes after ischemic stroke. Methods We followed the PRISMA guidelines for systematic reviews. Four databases (PubMed, Embase, Web of Science, and Cochrane Library) were searched comprehensively for studies published up to September 2025. Inclusion criteria encompassed original research focusing on ischemic stroke patients with mRS at 90 days as the primary outcome and predictive models utilizing imaging biomarkers from CT or MRI. Quality assessment of the included studies was conducted using the PROBAST and and METhodological RadiomICs Score (METRICS) tools. Results Seventeen studies were included in the quantitative meta-analysis. Models demonstrated good prognostic performance for 90-day outcomes, with pooled AUCs of 0.85 (95% CI: 0.80-0.86) for internal and 0.74 (95% CI: 0.67-0.89) for external validation sets, with no significant difference (p=0.42). Deep learning and hand-crafted radiomics models showed comparable diagnostic accuracy in the primary analysis (p=0.667); a difference in sensitivity favouring hand-crafted radiomics emerged only in a post-hoc analysis after exclusion of one outlier study. Models incorporating clinical variables significantly outperformed imaging-only models (p<0.01). CT- and MRI-based models showed similar performance. Treatment-specific analyses revealed higher sensitivity for IVT-based models and higher specificity for combined MT+IVT models. Conclusions DL and HCR models demonstrate robust potential for predicting 90-day outcomes in ischemic stroke. Future research should focus on standardizing model evaluation metrics and incorporating explainable AI techniques to improve clinical applicability.

Purpose To evaluate the external transportability of source-derived probability calibration for chest radiograph deep learning classifiers under distribution shift and determine its dependence on sample size and disease prevalence. Materials and Methods This retrospective secondary computational analysis of public, de-identified radiographic datasets involved no participant contact or new data collection. Cohort sizes were prespecified by availability without prospective power calculation: 106,686 NIH ChestX-ray14 images (29,323 patients; 1992-2015) for development and internal testing, 20,000 CheXpert images (20,000 patients; October 2002-July 2017), and 14,998 VinBigData images (2018-2020) for external testing. ResNet-18, DenseNet-121, and DeiT-Tiny classifiers for cardiomegaly and pleural effusion were trained across four patient fractions and three seeds (36 fits). Calibrators were fitted exclusively on NIH data. Twelve primary external Brier-score contrasts (temperature-scaled versus raw probabilities) were evaluated using paired cluster-bootstrapping with Holm adjustment. Results Temperature scaling reduced Brier scores in nine of 12 primary contrasts but worsened pleural effusion calibration on VinBigData across all architectures (Brier increases: ResNet-18, 0.001460 [95% CI: 0.001347, 0.001568]; DenseNet-121, 0.001193 [95% CI: 0.001109, 0.001274]; DeiT-Tiny, 0.000498 [95% CI: 0.000417, 0.000577]; all adjusted P=.006). Standardizing class weights to NIH prevalence reversed all CheXpert improvements. Although full versus 10% training improved external AUROC across all comparisons, DeiT-Tiny CheXpert pleural effusion Brier score worsened by 0.007039 despite a 0.032466 AUROC gain. At 80% selective image retention, calibrated ensembles retained only 12.3%-17.2% of VinBigData positive pleural effusion examinations. Conclusion Source-fitted calibration is not uniformly transportable across health systems, and discrimination gains do not guarantee calibrated probabilities. External evaluation must simultaneously assess probability reliability, prevalence shifts, and retained-case composition.

Head and neck radiation oncology, with its complex locality and epidemiology, requires various imaging techniques to achieve sufficient visualization. Through collaboration with medical physicists, radiation therapists, radiation oncologists, and neuroradiologists, our group clinically implemented a standard set of magnetic resonance imaging (MRI) sequences on an MR-simulation scanner for three anatomical regions: skull base, oral cavity, and larynx/pharynx. A total of 41 patients with head and neck cancer and 3 volunteers were scanned on either a 1.5T or 3T MR-simulation scanner. Sequences described include Dixon-based approaches, SPAIR-based approaches, short T2-based approaches, and quantitative imaging methods. We describe and outline the steps to optimize these notable MRI sequences to aid in successful implementation in a radiation oncology clinic.