Maria Romanovna Kodenko

Цель исследования: обоснование роли морфологических магнитно-резонансных (м-МР) признаков в гибридных морфо-радиомических моделях диагностики лимфоваскулярной инвазии (ЛВИ) злокачественного новообразования (ЗН) молочной железы (МЖ) и разработка прототипа объяснимой системы поддержки принятия врачебных решений (СППВР) на основе номограмм наивного Байеса с интегрированным SHAP-анализом. Материалы и методы. Нами анализировались данные 191 пациента: 13 м-МР признаков, 12390 радиомических (р-МР) показателей всего объёма опухоли (VOI ^entire ) и экстратуморальных тканей (VOI ^{+2 мм} ). Целевой признак — ЛВИ. Отбор предикторов выполнен дискретизацией методом Entropy-MDL; классификация — алгоритмом наивного Байеса с лапласовским сглаживанием ( α = 1 ). Прототип СППВР реализован как автономное single-page HTML-приложение с интегрированным расчётом логарифма отношения правдоподобия (log-LR) и SHAP-вкладов. Результаты. Получены две радиомические модели наивного Байеса (NB): NB_VOI ^entire (AUС=0,837) и NB_VOI ^{+2 мм} (AUС-ROC=0,887). По данным многомерной логистической регрессии перитуморальный отёк повышает шансы ЛВИ в 5,66 раз (OR 5,66, 95% ДИ: 2,27–14,94, pЗаключение. Прототип СППВР реализует концепцию объяснимой персонализированной диагностики посредством сочетания номограмм наивного Байеса и SHAP-анализа. Включение м-МР признаков перитуморального отёка и DWI-ободка в радиомические сигнатуры является не дополнительным, а структурообразующим элементом, обеспечивающим клиническую устойчивость диагностики.

Purpose of the study: to improve the reliability of prediction of lymphovascular invasion (LVI) by hybrid morpho-radiomic naive Bayesian models in patients with malignant breast cancer (MBC) by elucidating the role of morphologic magnetic resonance (m-MR) features. Materials and Methods. Data from 191 patients with MBC were analyzed in the form of 13 m-MR features, 6194 radiomic MR (r-MR) indicators of the whole tumor volume, and the target feature, LVI. Among the m-MR features, predictors of LVI were selected using crosstabulation, multivariate logistic regression, and Entropy-MDL discretization. Among 6194 r-MR indicators, predictors of LVI were selected by Entropy-MDL discretization. The selected indicators were used in training the naive Bayes algorithm. The performance of LVI predictors was compared. Results. According to multivariate logistic regression, the odds of LVI increased when tumor rim feature was detected on DWI image 4.05-fold (OR 4.05, 95%CI: 1.63–10.47, p = 0.003), peritumoral edema 5.66-fold (OR 5.66, 95%CI: 2.27–14.94, p < 0.001). 3 hybrid models with high discriminatory abilities were obtained: 1 model with DWI rim and radiomic signature from 4 p-MR indicators (AUC = 0.886, sensitivity – 89.5%, specificity – 79.1%, classification correctness – 89.5%, correctness of prediction of LVI – 73.3% and its absence – 95.2%), 2 model with peritumoral edema and radiomic signature from 6 p-MR indicators (AUC = 0.879, sensitivity – 82.5%, specificity – 80,9%, classification correctness – 82.5%, correctness in predicting LVI – 80.0% and its absence – 83.3%) and 3 models with peritumoral edema, rim DWI sign and radiomic signature from 8 p-MR indicators (AUS = 0.957, sensitivity – 96.5%, specificity – 90.2%, classification correctness – 96.5%, correctness in predicting LVI – 86.7% and its absence – 100%). Removing the DWI rim feature from 1 model worsens its discriminatory power (AUC-ROC, 0.801 ± 0.074 vs 0.886 ± 0.059, p = 0.001) and correctness of LVI prediction (40 vs 73%, p = 0.066). Similar but less pronounced, non-statistically significant changes were observed after removal of the peritumoral edema feature from the 2 models (AUC-ROC, 0.843 ± 0.067 vs 0.879 ± 0.060, p = 0.190; LVI prediction correctness, 60 vs 80%, p = 0.232). Removing 2 m-MR features from the 3 model worsens its discriminatory power (AUC-ROC, 0.957 ± 0.038 vs 0.901 ± 0.055, p = 0.024) and correctness of LVI prediction (80 vs 86.7%, p = 0.624). Conclusion. The use of hybrid models combining m-MR traits and r-MR indices improve the discriminatory power of prediction compared to models using only intratumoral r-MR indices.

Introduction . In the 21st century, medical imaging evolved from simple assessment of the presence of a pathological sign to quantitative assessment of detected changes. Despite the obvious dependence of diagnostic accuracy on the degree of measurement error, modern classifications of radiologic errors do not single out measurement error as a separate cause of discrepancies between specialists. We consider it necessary to study in detail the size of measurement error of radiologists in conditions close to routine practice and to determine the possibility of the influence of such an error on the formation of the correct patient management tactics. Aim . To determine radiologist discrepancy in measuring the pulmonary nodule size during opportunistic screening in chest computed tomography (CT). Materials and methods . A retrospective study was performed on a database containing the 3897 labeled native chest CT studies using a routine protocol. Experts determined the type and mean size of each pulmonary nodule. We used the following statistical methods: intraclass correlation coefficient to evaluate measurement reliability; limits of agreement (Bland-Altman method) to quantify measurement variability. The significance level for all statistical hypotheses was set to 0.05. Results . The final analysis included 87 CT studies, each containing a single solid pulmonary nodule with a diameter ranging from 6 to 30 mm, independently measured by three radiologists. The median nodule size was 10 mm. The intraclass correlation coefficient was 0.95 with 95% CI (0.93; 0.97). The 95% CI for the maximum absolute difference between observers ranged between (1.5; 1.9) mm. Conclusion . The measurement disperancy between three radiologists for solid pulmonary nodule in chest CT did not exceed the methodologically established threshold for pulmonary nodule growth.

Background/Objectives: Large language models (LLMs) are increasingly considered for use in radiology, including the summarization of patient medical records to support radiologists in processing large volumes of data under time constraints. This task requires not only accuracy and completeness but also clinical applicability. Automatic metrics and general-purpose questionnaires fail to capture these dimensions, and no standardized tool currently exists for the expert evaluation of LLM-generated summaries in radiology. Here, we aimed to develop and validate such a tool. Methods: Items for the questionnaire were formulated and refined through focus group testing with radiologists. Validation was performed on 132 LLM-generated summaries of 44 patient records, each independently assessed by radiologists. Criterion validity was evaluated through known-group differentiation and construct validity through confirmatory factor analysis. Results: The resulting seven-item instrument, ELEGANCE (Expert-Led Evaluation of Generative AI Competence and Excellence), demonstrated excellent internal consistency (Cronbach’s α = 0.95). It encompasses seven dimensions: relevance, completeness, applicability, falsification, satisfaction, structure, and correctness of language and terminology. Confirmatory factor analysis supported a two-factor structure (content and form), with strong fit indices (RMSEA = 0.079, CFI = 0.989, TLI = 0.982, SRMR = 0.029). Criterion validity was confirmed by significant between-group differences (p < 0.001). Conclusions: ELEGANCE is the first validated tool for expert evaluation of LLM-generated medical record summaries for radiologists, providing a standardized framework to ensure quality and clinical utility.

BACKGROUND: Texture analysis improves the diagnostic accuracy of magnetic resonance imaging and differential diagnosis of prostate lesions, which are primarily segmented through manual labeling, resulting in significant inter-expert variability of masks. A consensus-based technique can help reduce inconsistencies in prostate lesion segmentation. However, global scientific studies have not described any standardized, consensus-based labeling protocols. AIM: This study aimed to develop a consensus algorithm for manual labeling of prostate lesions by several independent experts and evaluate inter-expert consistency in lesion segmentation. METHODS: This retrospective study included 60 biparametric magnetic resonance imaging scans of the prostate gland performed according to PI-RADS 2.1 technical specification. The scans showed PI-RADS 3, 4, and 5 lesions. Two independent radiologists manually segmented the prostate lesions using 3D Slicer. Then, the resulting masks were compared using the Dice–Sørensen coefficient (DSC). For lesions with DSC ≥ 0.75, the final mask was based on the overlap between the two original masks. Conversely, for lesions with DSC 0.75, the final mask was determined using the proposed consensus algorithm. RESULTS: The proposed consensus algorithm significantly increased the DSC values, from 0.61 [0.48; 0.73] for primary labeling to 0.74 [0.62; 0.79] for labeling using the proposed algorithm (p = 0.01). CONCLUSION: The proposed consensus-based algorithm for labeling prostate lesions using magnetic resonance imaging data is crucial in addressing inadequate approaches to objective segmentation in research and clinical settings.

Objective To develop a platform for automated processing of radiological datasets that operates independently of medical information systems. The platform maintains datasets throughout their lifecycle, from data retrieval to annotation and presentation. Methods The platform employs a modular structure in which modules can operate independently or in conjunction. Each module sequentially processes output from the preceding module. The platform incorporates a local database containing textual study protocols, a radiology information system (RIS), and storage for labeled studies and reports. Results A platform equipped with local permanent and temporary file storages facilitates radiological datasets processing. The platform’s modules enable data search, extraction, anonymization, annotation, generation of annotated files, and standardized documentation of datasets. Conclusion The platform provides a comprehensive workflow for radiological dataset management and is currently operational at the Center for Diagnostics and Telemedicine. Future development will focus on expanding platform functionality.

The accurate assessment of muscle morphology and function is crucial for medical diagnostics, rehabilitation, and biomechanical research. This study presents a novel methodology for constructing volumetric models of forearm muscles based on three-dimensional ultrasound imaging integrated with a robotic system to ensure precise probe positioning and controlled pressure application. The proposed ultrasound scanning approach combined with a collaborative six-degrees-of-freedom robotic manipulator enabled reproducible and high-resolution imaging of muscle structures in both relaxed and contracted states. A custom-built phantom, acoustically similar to biological tissues, was developed to validate the method. The cross-sectional area of the muscles and the coordinates of the center of mass of the sections, as well as the volume and center of gravity of each muscle, were calculated for each cross-section of the reconstructed forearm muscle models at contraction. The method’s feasibility was confirmed by comparing the reconstructed volumes with anatomical data and phantom measurements. This study highlights the advantages of robotic-assisted ultrasound imaging for non-invasive muscle assessment and suggests its potential applications in neuromuscular diagnostics, prosthetics design, and rehabilitation monitoring.

BACKGROUND: Artificial intelligence algorithms are used to analyze images obtained through radiological diagnostic methods. The effectiveness of such algorithms depends on the availability of relevant and representative training datasets. The volume of such data in the public domain should be increased, particularly datasets containing abdominal aorta computed tomography angiography images, with pathology classification and vessel segmentation. The limitations of existing solutions include small sample sizes, restricted dataset specialization, and inconsistent dataset preparation methodologies. Aim: To create an open dataset containing computed tomography angiography images of abdominal aorta segmentation for normal aorta, aneurysm, thrombosis, and calcification. MATERIALS AND METHODS: A technical specification for dataset preparation was developed according to the methodology for testing artificial intelligence algorithms, the required sample size was calculated, and approval was obtained from an independent ethics committee. Regarding dataset creation, a previously developed original semiautomatic segmentation algorithm using Slicer 3D software was employed. The inclusion criteria were computed tomography angiography or abdominal computed tomography scans with contrast, arterial phase, and slice thickness ≤3 mm. Conversely, the exclusion criteria were presence of foreign bodies in the aorta lumen and aortic dissection. The algorithm was tested on patient data obtained from the Unified Radiological Information System. An expert evaluation was conducted to assess the compliance of obtained results with the established requirements and evaluate the time efficiency of using the developed segmentation algorithm. RESULTS: The calculated sample size was 100 angiographic studies, including arterial phase scans with a slice thickness of ≤1.2 mm. Population data: number of unique patients, 100; percentage of female patients, 51%; and median age, 62 years (age range: 18–84 years). Pathology (including combined pathology) was detected in 61% of cases: 60 studies showed signs of calcification, 18 revealed aortic dilation, and 18 determined signs of thrombosed lumen. The average time to process one study (100 slices) using the developed segmentation algorithm was 0.8 hours. CONCLUSIONS: A dataset containing 100 computed tomography angiography results with abdominal aorta segmentation for normal cases, aneurysm, thrombosis, and calcification was created. The dataset is publicly available and can be used for developing and testing artificial intelligence algorithms and for anthropomorphic modeling of the abdominal aorta.