Diagnostic performance and short-interval reproducibility of multimodal large language models in differentiating cholesteatoma from chronic otitis media using key-image temporal bone high-resolution computed tomography.
Aug 2026· Diagnostic and Interventional Radiology· 0 citations
Medicine
TL;DR
The evaluated models lacked the spatial precision and consistency required for the accurate assessment of complex middle ear structures and underscores the necessity for verification by a radiologist and continuous monitoring.
Abstract
Purpose
This study aimed to assess whether multimodal large language models (LLMs) can distinguish cholesteatoma from non-cholesteatomatous chronic otitis media (COM) on representative key-image temporal bone high-resolution computed tomography (HRCT) and to evaluate the short-interval reproducibility of their outputs.
Methods
This retrospective, single-center study (2019-2024) included 101 patients (48 with cholesteatoma, 53 with non-cholesteatomatous COM) who underwent surgical treatment. The reference standard was intraoperative diagnosis with histopathological confirmation for cholesteatoma and surgical documentation for COM. For each case, six anonymized representative HRCT images reflecting standard diagnostic criteria were selected by consensus between a 4th-year radiology resident and a board-certified head and neck radiologist, both unaware of the diagnosis. Subsequently, the same images were analyzed by the GPT-5 and Gemini 2.5 Pro LLMs through their official web interfaces, utilizing structured prompts and a zero-shot approach. These evaluations were conducted in two distinct sessions (S1 and S2) with a 1-week interval. The primary endpoint was accurate binary classification. Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated with 95% confidence intervals [(CIs); Wilson method] agreement with the reference standard and between sessions and models was assessed with the Cohen kappa (κ) coefficient; and differences in classification were assessed with the McNemar test.
Results
The radiologist achieved an accuracy of 96.0% (95% CI: 90.3-98.4) with almost perfect agreement with the reference standard (κ: 0.921). In S1, GPT-5 and Gemini 2.5 Pro achieved accuracies of 43.6% and 49.5%, and in S2, 46.5% and 47.5%, respectively. Both models combined high sensitivity (83.3%-97.9%) with low specificity (1.9%-13.2%), and balanced accuracy ranged from 0.45 to 0.52. Between-session reproducibility was fair for GPT-5 (κ: 0.360) and moderate for Gemini 2.5 Pro (κ: 0.485), and inter-model agreement was slight at both sessions (κ: 0.035 at S1 and κ: 0.086 at S2). Accuracy did not differ significantly between the two models (P = 0.211).
Conclusion
In this single-center study, GPT-5 and Gemini 2.5 Pro, in the versions evaluated, combined high sensitivity with low specificity and showed only fair-to-moderate between-session reproducibility and exhibited slight inter-model agreement on temporal bone key-image HRCT. These findings do not support their use as independent second readers, and broader generalization to other multimodal LLMs would require the evaluation of additional models.
CLINICAL
Significance
The evaluated models lacked the spatial precision and consistency required for the accurate assessment of complex middle ear structures. This finding underscores the necessity for verification by a radiologist and continuous monitoring.
OBJECTIVE
To compare the diagnostic performance of high-resolution computed tomography (HRCT) and non-echo-planar diffusion-weighted magnetic resonance imaging (non-EPI DWI-MRI) for middle ear cholesteatoma and to assess the added value of their combined use in preoperative evaluation.
METHODS
Imaging data sets from 85 subjects (65 with surgically or histopathologically confirmed cholesteatoma, 20 controls) were retrospectively analyzed. HRCT, non-EPI DWI-MRI, and combined HRCT+DWI-MRI interpretations were compared against reference standards. Diagnostic metrics included sensitivity, specificity, accuracy, localization error, and interobserver agreement (κ).
RESULTS
HRCT demonstrated high sensitivity (95.4%) for detecting bony erosions and ossicular abnormalities, whereas non-EPI DWI-MRI showed greater specificity (85.0%) in distinguishing cholesteatoma from other soft tissues. The combined approach yielded optimal diagnostic performance (sensitivity 93.8%, specificity 92.0%, accuracy 93.2%) and the lowest localization error (mean 1.8 mm). Interobserver agreement was excellent (κ=0.88).
CONCLUSION
High-resolution computed tomography and non-echo-planar diffusion-weighted MRI provide complementary diagnostic information in the evaluation of middle ear cholesteatoma. Their combined interpretation may improve diagnostic confidence and radiologic assessment of lesion extent; however, the direct impact on surgical outcomes requires further prospective validation.
Huimin Xu, Gao Gang, Xinhong Lu et al.· Journal of computer assisted...· 0 citations
Cholesteatoma is a benign but locally destructive lesion of the middle ear and mastoid treated surgically, with a well-recognized risk of postoperative recurrence. This systematic review evaluated the diagnostic performance of non-echo-planar diffusion-weighted magnetic resonance imaging (non-EPI DWI) in detecting postoperative recurrences and assessed variability in follow-up protocols. A systematic search was conducted in Embase, Scopus, PubMed, CINAHL, the Cochrane Library, and SciELO for studies published between 2000 and November 2025, following PRISMA guidelines. Eligible studies included postoperative cholesteatoma patients undergoing follow-up with non-EPI DWI and reporting diagnostic performance and/or follow-up characteristics. Out of the 115 identified articles, 9 studies involving 931 patients met the inclusion criteria. Non-EPI DWI demonstrated high specificity (90–100%) and variable sensitivity (59–100%). Positive predictive values ranged from 93 to 100%, while negative predictive values ranged from 53 to 91.7%. Detection of cholesteatomas smaller than 2–3 mm was limited across studies. Follow-up duration varied widely from 1 to 31.6 years, and recurrences were identified both early and late, up to 7.5 years postoperatively. Marked heterogeneity was observed in imaging protocols and follow-up schedules. Non-EPI DWI demonstrates consistently high specificity and variable sensitivity for detecting postoperative cholesteatoma recurrence, particularly for lesions larger than 2–3 mm. However, substantial heterogeneity across studies limits definitive conclusions regarding optimal surveillance strategies. These findings highlight the need for standardized imaging protocols and long-term follow-up tailored to surgical technique and patient risk profile.
Dolores Rodríguez Rodríguez, Cristina Isabel Sanz Sánchez, Eva Flores Carmona et al.· The Egyptian Journal of Radi...· 0 citations
The findings support the potential role of MLLMs as assistive tools in human-AI orthopedic imaging workflows, but external validation, careful input standardization, and prospective clinical evaluation are needed before clinical deployment.
Jiesheng Zhu, Xingxing Huang, Jincheng Shi et al.· Journal of Medical Internet...· 0 citations
Objective: To determine the diagnostic accuracy of unenhanced low-dose multi-detector computed tomography (LDMDCT) in detecting acute appendicitis, taking histopathology as the gold standard.
Study Design: Cross-sectional validation study.
Place and Duration of Study: Department of Radiology, Combined Military Hospital Rawalpindi, Pakistan, from May to Nov 2025.
Methodology: A total of 92 patients aged 18–40 years presenting with clinically suspected Acute Appendicitis were enrolled using non-probability consecutive sampling. All patients underwent unenhanced LDMDCT on a GE Optima 128-slice scanner with radiation dose maintained below 3 mSv. CT findings were evaluated against histopathological diagnosis of resected appendix specimens. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall diagnostic accuracy were calculated.
Results: The mean age was 27.4±6.8 years, and 58(63.0%) participants were male. Histopathology confirmed acute appendicitis in 51 (55.4%) patients. Unenhanced LDMDCT demonstrated sensitivity of 88.2% (95% CI: 76.1–95.6%), specificity of 87.8% (95% CI: 73.8–95.9%), PPV of 90.0%, NPV of 85.7%, and overall diagnostic accuracy of 88.0%. No statistically significant difference in diagnostic accuracy was observed across gender or age subgroups (p>0.05).
Conclusion: Unenhanced LDMDCT is a reliable and radiation-conscious imaging modality with high diagnostic accuracy for suspected acute appendicitis. Its routine adoption in emergency imaging pathways is strongly recommended to reduce unnecessary surgery and curtail radiation burden in the predominantly young population affected by this condition.
Introduction: Focal liver lesions (FLLs) are common findings in clinical practice, requiring accurate characterization to guide patient management. Multidetector computed tomography (MDCT) is widely available and routinely used in Bhutan. Therefore, this study aimed to determine the diagnostic accuracy of multidetector computed tomography (MDCT) in detecting focal liver lesions compared with histopathological findings.
Methods: This cross-sectional study was conducted at the Radiology Department of the Jigme Dorji Wangchuck National Referral Hospital from January 2024 to December 2025. All patients who underwent MDCT were included in the study. Data were collected by reviewing medical documents and recorded using a standard pro forma, and data were analyzed using SPSS (version 26).
Results: The majority were female (57.1%), with a mean age of 57.49 years. Histopathological examination revealed 79.4% malignant and 20.6% benign lesions. MDCT correctly characterized 49 malignant lesions (72%) and 9 benign lesions (8%), yielding an overall diagnostic accuracy of 92.1%. The sensitivity and specificity of MDCT for differentiating malignant from benign lesions were 98.0% and 69.2%, respectively, with a PPV of 92.5% and NPV of 90.0%. Cohen's kappa coefficient showed substantial agreement between MDCT and histopathology (kappa = 0.74). The association between MDCT diagnosis and histopathology was statistically significant (chi-square = 34.923, p < 0.001).
Conclusion: MDCT demonstrates high diagnostic accuracy in characterizing focal liver lesions, with excellent sensitivity for detecting malignant lesions. These findings support the continued use of MDCT as the primary imaging modality in Bhutan, where access to advanced imaging modalities is limited.
Tashi Gyeltshen, Gyem Dorji, Kanokwan Wetasin et al.· Bhutan health journal· 0 citations
Background Accurate preoperative characterization of solitary pulmonary nodules (SPNs), particularly differentiation of benign from malignant lesions and assessment of adenocarcinoma invasiveness, remains challenging based on morphological features alone. We investigated whether integrating 1024-matrix ultra-high-resolution computed tomography (UHRCT) with two liquid-biopsy signals, urinary cellular energy metabolism (CEM) and a seven tumor-associated autoantibody (7-TAAb) panel, could improve diagnostic performance in surgically treated patients with SPNs. Methods In this single-center prospective study, consecutive adults with SPNs (≤3 cm) scheduled for video-assisted thoracoscopic surgery between January 2022 and December 2024 were enrolled (n = 183). All patients underwent preoperative UHRCT (1024×1024), urinary CEM testing, and 7-TAAb testing. A paired subgroup (n = 85) additionally underwent conventional HRCT (512×512) within 1 month before surgery. Postoperative histopathology served as the reference standard. Diagnostic performance metrics were calculated for each modality, parallel-rule combined strategies, and probability-based logistic regression models. Results Among 183 patients, 164 (89.6%) had malignant nodules. UHRCT achieved the best single-modality discrimination (AUC, 0.831; sensitivity, 87.2%; specificity, 78.9%), outperforming CEM (AUC, 0.605) and 7-TAAbs (AUC, 0.558). Under the parallel-rule strategy, multimodal testing increased sensitivity but markedly reduced specificity (sensitivity 98.8%, specificity 9.5%). The probability-based fully integrated model achieved the highest AUC (0.946). At the Youden-index threshold of 0.816, this model yielded sensitivity of 92.7%, specificity of 84.2%, PPV of 98.1%, NPV of 57.1%, and accuracy of 91.8%. In the paired comparison (n = 85), UHRCT showed higher sensitivity (92.0% vs. 68.0%) and specificity (90.0% vs. 73.3%) than HRCT. For invasiveness assessment, positivity rates of both urinary CEM (57.6% vs. 31.6%) and 7-TAAbs (53.8% vs. 31.0%) were higher in invasive adenocarcinoma than in pre-invasive lesions. Conclusion In surgically selected patients with SPNs, 1024-matrix UHRCT provided the strongest single-modality diagnostic information, while urinary CEM and 7-TAAbs offered complementary biological signals associated with invasive adenocarcinoma pathology. Probability-based integration provided a more balanced diagnostic framework than a simple parallel-rule strategy. Given the high prevalence of malignancy and the limited number of benign lesions, these findings should be considered exploratory and require external validation in broader outpatient and screening cohorts.
Zhengjun Li, Siwei Chao, Ding Li et al.· Frontiers in Oncology· 0 citations