European Radiology Experimental最新文献

Background: Fetal MRI is increasingly used to investigate fetal lung pathologies, and super-resolution (SR) algorithms could be a powerful clinical tool for this assessment. Our goal was to investigate whether SR reconstructions result in an improved agreement in lung volume measurements determined by different raters, also known as inter-rater reliability.

Materials and methods: In this single-center retrospective study, fetal lung volumes calculated from both SR reconstructions and the original images were analyzed. Three radiologists manually segmented the fetal lungs and rated the image quality of all images and reconstructions. Fetal lung volumes were calculated, and the coefficient of variation (CV) was determined for each set of images. Bland-Altman plots were generated, and intraclass correlation coefficients (ICCs) were calculated. A one-sided paired Wilcoxon test was used to compare the fetal lung volume CVs, and a two-sided paired t-test was used to compare the lung volumes. The quality ratings were compared using a two-sided paired Wilcoxon test.

Results: A total of 98 fetal scans with gestational ages from 19 to 37 weeks were evaluated. There was a significantly lower CV in the lung volumes segmented from the SR reconstructions (p < 0.001), and the ICCs of the reconstructions were higher than those determined from the original images. Bland-Altman plots demonstrated better agreement in the SR reconstruction lung volumes. No significant differences in quality ratings or lung volumes were found.

Conclusion: SR reconstructions of the fetal lungs in MRI enabled better inter-rater reliability of fetal lung volume assessment.

Relevance statement: SR reconstructions of the fetal body, obtained through fetal MRI, can be a valuable tool for improving the inter-rater reliability of fetal lung volume measurements, a crucial clinical biomarker for assessing fetal development and predicting pregnancy outcomes.

Key points: Deformable slice-to-volume reconstructions of fetal body MRI could be a valuable clinical tool. Quantitative advantages of fetal body MRI reconstructions need to be proven. Fetal body MRI reconstructions improved inter-rater reliability in lung volume measurements.

Background: Lung lobe segmentation is required to assess lobar function with nuclear imaging before surgical interventions. We evaluated the performance of open-source deep learning-based lung lobe segmentation tools, compared to a similar nnU-Net model trained on a smaller but more representative clinical dataset.

Materials and methods: We collated and semi-automatically segmented an internal dataset of 164 computed tomography scans and classified them for task difficulty as easy, moderate, or hard. The performance of three open-source models-multi-organ objective segmentation (MOOSE), TotalSegmentator, and LungMask-was assessed using Dice similarity coefficient (DSC), robust Hausdorff distance (rHd95), and normalized surface distance (NSD). Additionally, we trained, validated, and tested an nnU-Net model using our local dataset and compared its performance with that of the other software on the test subset. All models were evaluated for generalizability using an external competition (LOLA11, n = 55).

Results: TotalSegmentator outperformed MOOSE in DSC and NSD across all difficulty levels (p < 0.001), but not in rHd95 (p = 1.000). MOOSE and TotalSegmentator surpassed LungMask across metrics and difficulty classes (p < 0.001). Our model exceeded all other models on the internal dataset (n = 33) in all metrics, across all difficulty classes (p < 0.001), and on the external dataset. Missing lobes were correctly identified only by our model and LungMask in 3 and 1 of 7 cases, respectively.

Conclusion: Open-source segmentation tools perform well in straightforward cases but struggle in unfamiliar, complex cases. Training on diverse, specialized datasets can improve generalizability, emphasizing representative data over sheer quantity.

Relevance statement: Training lung lobe segmentation models on a local variety of cases improves accuracy, thus enhancing presurgical planning, ventilation-perfusion analysis, and disease localization, potentially impacting treatment decisions and patient outcomes in respiratory and thoracic care.

Key points: Deep learning models trained on non-specialized datasets struggle with complex lung anomalies, yet their real-world limitations are insufficiently assessed. Training an identical model on a smaller yet clinically diverse and representative cohort improved performance in challenging cases. Data diversity outweighs the quantity in deep learning-based segmentation models. Accurate lung lobe segmentation may enhance presurgical assessment of lung lobar ventilation and perfusion function, optimizing clinical decision-making and patient outcomes.

In radiomics, features are often linked to biomarkers and are generally expected to be reproducible, as reproducibility is considered a prerequisite for developing predictive models in clinical applications. However, this perspective overlooks feature interactions and may underestimate the potential value of nonreproducible features. Through experiments simulating a test-retest scenario, we demonstrate that even non-reproducible features can contribute significantly to predictive performance. Removing these features can lower model accuracy. These findings suggest that the emphasis on feature reproducibility should be reconsidered and that features should not be evaluated in isolation. Underlying information can be spread across multiple features. Focusing on individual features ignores feature interactions and may limit the model's predictive power. Ultimately, radiomics must prioritize prediction and clinical relevance. KEY POINTS: Feature reproducibility assessments often ignore feature interactions, overlooking predictive performance. Feature reproducibility depends on subjective thresholds, chosen metrics, and sample size. Nonreproducible features can be more predictive than reproducible ones. Predictive information may be distributed across multiple features rather than confined to individual ones.

Background: Electrochemotherapy (ECT) of vertebral metastasis is a new treatment option for metastasis that is not accessible to thermal ablation or radiotherapy. A numerical feasibility study has investigated the transpedicular approach for electrode insertion. We conducted a preclinical study to assess its safety.

Methods: Histologic examination of the spinal cord was performed in 12 consecutive pigs treated with ECT at three consecutive levels (T11, T12, and L1) to evaluate any cellular or vascular damage. Pigs of group A (n = 6) had an intraoperative neuromonitoring immediately for 1 h after ECT and then were euthanized. Pain and clinical symptoms were daily evaluated for group B (n = 3) and group C (n = 3) until day-3 and day-30, respectively.

Results: At gross pathology, no apoptosis, no vascular/thrombosis or hemorrhagic focus was observed in any pig. Motor-evoked potential responses of the lower limbs were transiently lost in response in 5 of the 6 pigs, but complete recovery always occurred within 30 min. Clinical examination (groups B and C) revealed no symptoms during the follow-up. Pigs were all able to walk normally, without weakness or paralysis of the lower extremities. No urinary/fecal retention or incontinence was observed, nor any sign of pain.

Conclusion: Our results confirm that the insertion of electrodes through the pedicles is safe for the ECT of vertebral metastases. Further studies are needed to evaluate the safety profile of ECT of vertebral metastases invading the cortical and epidural fat, which represents a privileged pathway for the electric field between the electrodes.

Relevance statement: Electrochemotherapy of vertebral metastases should be performed using a transpedicular approach for the insertion of electrodes, without definitive sequelae at the spinal cord level.

Key points: Electrochemotherapy is a new treatment for vertebral metastases not accessible to radiotherapy, but it could result in spinal cord injury related to electrical trauma. In a swine model, the transpedicular approach has demonstrated no definitive sequelae at intraoperative neuromonitoring and during clinical follow-up. Electrochemotherapy should be performed using a transpedicular approach to avoid spinal cord damage.

Background: The value of virtual noncontrast (VNC) images from photon-counting computed tomography (PCCT) for evaluating adrenal lesions and diagnosing adrenal adenomas remains to be clarified.

Materials and methods: Participants with adrenal masses who underwent unenhanced and portal venous phase PCCT were prospectively included. Portal-venous phase images were reconstructed using conventional VNC (VNC_Conv) and PureCalcium VNC (VNC_PC). We measured two-dimensional (2D) attenuation of adrenal masses at their largest slice on true noncontrast (TNC), VNC_Conv, and VNC_PC images. Three-dimensional (3D) attenuation and radiomic features of adrenal masses were semiautomatically extracted. These parameters were statistically compared, and diagnostic performance for adenomas was evaluated.

Results: The study included 54 participants (27 females, mean age 45.3 years) with 68 adrenal lesions. Attenuation values on VNC were higher than those on TNC. TNC, VNC_Conv, and VNC_PC attenuation values did not differ between 2D and 3D measurements. The intraclass correlation coefficients of first-order, shape, and texture features between TNC and VNC were 0.671, 0.822, and 0.616, respectively. The sensitivity and specificity of the proposed thresholds (VNC_Conv 25 HU, VNC_PC 20 HU) were higher than those of the previously established threshold of 10 HU in diagnosing adenomas. There was no significant difference between VNC_Conv and VNC_PC in diagnosing adenomas (area under the receiver operating characteristic curve: 0.841 versus 0.838, p = 0.873).

Conclusion: VNC algorithms from PCCT overestimated CT attenuation of adrenal lesions. Higher thresholds showed better diagnostic performance for discriminating adrenal adenomas from non-adenomas than the established 10 HU.

Relevance statement: We investigated the application of VNC images from PCCT in adrenal disease. On VNC images, higher thresholds, superior to the accepted 10 HU, are needed for discriminating adenomas from non-adenomas, reducing the need for secondary examinations.

Key points: This study investigated the value of VNC images from PCCT in adrenal lesions. VNC reconstruction overestimated the CT attenuation of adrenal lesions. Higher thresholds on VNC images were superior to the accepted 10 HU for differentiating adenomas from non-adenomas.

Background: This study aimed to evaluate the imaging performance and diagnostic value of a photon-counting detector (PCD) computed tomography (CT) compared to an energy-integrating detector (EID) and flat panel detector (FPD) for cadaveric wrist arthrographies.

Methods: Following ethics committee approval, ten cadaveric wrists were injected with diluted iodinated contrast agent. CT arthrographies using PCD-, EID-, and FPD-CT were performed. Six dose protocols between 0.1 mGy (using a tin filter) and 6 mGy, ultrahigh-resolution-mode, and two reconstruction kernels were used for the PCD-CT and EID-CT. FPD-CT images were reconstructed using a "normal" and "sharp" kernel. Signal-to-noise ratios (SNR) and contrast-to-noise ratios (CNR) were calculated and analyzed using analysis of variance (ANOVA) and post hoc tests. Three blinded radiologists independently rated image quality concerning trabecular, cartilage, and intrinsic structures. Intraclass correlation coefficients (ICC) were calculated, followed by a Friedman and post hoc test.

Results: At 1.5 mGy, 3 mGy, and 6 mGy with the Br89 kernel, the PCD-CT yielded up to 2.35 times higher SNR and up to 7 times higher CNR than dose-equivalent and higher dose EID-CT scans. Subjective ratings favored the PCD-CT over the EID-CT and occasionally the FPD-CT, with a combined ICC of 0.942. Applying sharper kernels, SNR did not differ significantly between the PCD-CT (1.5 mGy, 3 mGy, and 6 mGy) and the FPD-CT.

Conclusion: Using sharp kernels, the PCD-CT provided superior image quality to the EID-CT and achieved comparable or better quality than the FPD at certain parameters. Thus, the PCD-CT could be considered a possible alternative in clinical routine for evaluating wrist injuries.

Relevance statement: This study demonstrates the potential of the PCD-CT as a valuable tool in diagnosing wrist injuries. Its superior image quality compared to the EID-CT can increase confidence in diagnosing subtle bone pathologies and additionally yields the possibility of radiation exposure reduction.

Key points: The technical advantages of the PCD-CT allow for dose reduction while generating high-quality images. PCD-CT showed superior image quality over EID-CT and was comparable to the FPD-CT. PDC-CT offers improved visualization of fine joint structures in wrist arthrography and should be considered in clinical routine.

Background: The onset of liver-related events (LREs) in fibrosis indicates a poor prognosis and worsens patients' quality of life, making the prediction and early detection of LREs crucial. The aim of this study was to develop a radiomics model using liver magnetic resonance imaging (MRI) to predict LRE risk in patients undergoing antiviral treatment for chronic fibrosis caused by hepatitis B virus (HBV).

Methods: Patients with HBV-associated liver fibrosis and liver stiffness measurements ≥ 10 kPa were included. Feature selection and dimensionality reduction techniques identified discriminative features from three MRI sequences. Radiomics models were built using eight machine learning techniques and evaluated for performance. Shapley additive explanation and permutation importance techniques were applied to interpret the model output.

Results: A total of 222 patients aged 49 ± 10 years (mean ± standard deviation), 175 males, were evaluated, with 41 experiencing LREs. The radiomics model, incorporating 58 selected features, outperformed traditional clinical tools in prediction accuracy. Developed using a support vector machine classifier, the model achieved optimal areas under the receiver operating characteristic curves of 0.94 and 0.93 in the training and test sets, respectively, demonstrating good calibration.

Conclusion: Machine learning techniques effectively predicted LREs in patients with fibrosis and HBV, offering comparable accuracy across algorithms and supporting personalized care decisions for HBV-related liver disease.

Relevance statement: Radiomics models based on liver multisequence MRI can improve risk prediction and management of patients with HBV-associated chronic fibrosis. In addition, it offers valuable prognostic insights and aids in making informed clinical decisions.

Key points: Liver-related events (LREs) are associated with poor prognosis in chronic fibrosis. Radiomics models could predict LREs in patients with hepatitis B-associated chronic fibrosis. Radiomics contributes to personalized care choices for patients with hepatitis B-associated fibrosis.

Background: R2* and quantitative susceptibility mapping (QSM) are regarded as robust techniques for assessing iron content in the brain. While these techniques are established for normal or moderate iron levels, their usability in extreme iron overload, as seen in aceruloplasminemia (ACP), is unclear. We aimed to evaluate various R2* and QSM algorithms in assessing brain iron levels in patients with ACP compared to healthy controls.

Materials and methods: We acquired a three-dimensional multiecho gradient-echo sequence for R2* and QSM in three patients with ACP and three healthy subjects. Six algorithms each for R2* and QSM were compared. QSM was performed with referencing to whole brain, to cerebrospinal fluid and without referencing. R2* and QSM values were assessed in the caudate nucleus, putamen, globus pallidus, and thalamus.

Results: R2* values varied significantly across algorithms, particularly in the putamen (F(5,50) = 16.51, p < 0.001). For QSM, reference region choice (F(5,150) = 264, p < 0.001) and algorithm selection (F(2,9) = 10, p < 0.001) had an impact on susceptibility values. In patients, referencing to whole brain yielded lower susceptibility values than cerebrospinal fluid (median = 0.147 ppm, range = 0.527 ppm versus median = 0.279 ppm, range = 0.593 ppm).

Conclusion: Extreme iron overload amplifies variability in R2* and QSM measurements. QSM referencing is particularly challenging in diffuse whole-brain iron accumulation; thus, analysis with multiple reference regions might mitigate bias. Both algorithm selection and referencing approaches play a pivotal role in determining measurement accuracy and clinical interpretation under extreme brain iron overload.

Relevance statement: As QSM transitions into clinical use, it will encounter cases of extreme iron overload. Our study in patients with aceruloplasminemia revealed that the choice of reference region significantly influences susceptibility values, with variations exceeding algorithm-dependent differences.

Key points: R2* and QSM vary across algorithms in brain tissue with iron overload. Whole-brain referenced QSM leads to lower susceptibility values in aceruloplasminemia patients. QSM, if properly processed, provides reliable maps in iron overload brain regions. In brain regions with extremely high iron content, R2* mapping might fail.

Background: We compared the performance, confidence, and response consistency of five chatbots powered by large language models in solving European Diploma in Radiology (EDiR) text-based multiple-response questions.

Methods: ChatGPT-4o, ChatGPT-4o-mini, Copilot, Gemini, and Claude 3.5 Sonnet were tested using 52 text-based multiple-response questions from two previous EDiR sessions in two iterations. Chatbots were prompted to evaluate each answer as correct or incorrect and grade its confidence level on a scale of 0 (not confident at all) to 10 (most confident). Scores per question were calculated using a weighted formula that accounted for correct and incorrect answers (range 0.0-1.0).

Results: Claude 3.5 Sonnet achieved the highest score per question (0.84 ± 0.26, mean ± standard deviation) compared to ChatGPT-4o (0.76 ± 0.31), ChatGPT-4o-mini (0.64 ± 0.35), Copilot (0.62 ± 0.37), and Gemini (0.54 ± 0.39) (p < 0.001). A self-reported confidence in answering the questions was 9.0 ± 0.9 for Claude 3.5 Sonnet followed by ChatGPT-4o (8.7 ± 1.1), compared to ChatGPT-4o-mini (8.2 ± 1.3), Copilot (8.2 ± 2.2), and Gemini (8.2 ± 1.6, p < 0.001). Claude 3.5 Sonnet demonstrated superior consistency, changing responses in 5.4% of cases between the two iterations, compared to ChatGPT-4o (6.5%), ChatGPT-4o-mini (8.8%), Copilot (13.8%), and Gemini (18.5%). All chatbots outperformed human candidates from previous EDiR sessions, achieving a passing grade from this part of the examination.

Conclusion: Claude 3.5 Sonnet exhibited superior accuracy, confidence, and consistency, with ChatGPT-4o performing nearly as well. The variation in performance among the evaluated models was substantial.

Relevance statement: Variation in performance, consistency, and confidence among chatbots in solving EDiR test-based questions highlights the need for cautious deployment, particularly in high-stakes clinical and educational settings.

Key points: Claude 3.5 Sonnet outperformed other chatbots in accuracy and response consistency. ChatGPT-4o ranked second, showing strong but slightly less reliable performance. All chatbots surpassed EDiR candidates in text-based EDiR questions.

Background: Accurate segmentation of lung cancer lesions in computed tomography (CT) is essential for precise diagnosis, personalized therapy planning, and treatment response assessment. While automatic segmentation of the primary lung lesion has been widely studied, the ability to segment multiple lesions per patient remains underexplored. In this study, we address this gap by introducing a novel, automated approach for multi-instance segmentation of lung cancer lesions, leveraging a heterogeneous cohort with real-world multicenter data.

Materials and methods: We analyzed 1,081 retrospectively collected CT scans with 5,322 annotated lesions (4.92 ± 13.05 lesions per scan). The cohort was stratified into training (n = 868) and testing (n = 213) subsets. We developed an automated three-step pipeline, including thoracic bounding box extraction, multi-instance lesion segmentation, and false positive reduction via a novel multiscale cascade classifier to filter spurious and non-lesion candidates.

Results: On the independent test set, our method achieved a Dice similarity coefficient of 76% for segmentation and a lesion detection sensitivity of 85%. When evaluated on an external dataset of 188 real-world cases, it achieved a Dice similarity coefficient of 73%, and a lesion detection sensitivity of 85%.

Conclusion: Our approach accurately detected and segmented multiple lung cancer lesions per patient on CT scans, demonstrating robustness across an independent test set and an external real-world dataset.

Relevance statement: AI-driven segmentation comprehensively captures lesion burden, enhancing lung cancer assessment and disease monitoring KEY POINTS: Automatic multi-instance lung cancer lesion segmentation is underexplored yet crucial for disease assessment. Developed a deep learning-based segmentation pipeline trained on multi-center real-world data, which reached 85% sensitivity at external validation. Thoracic bounding box and false positive reduction techniques improved the pipeline's segmentation performance.