Background Large language models (LLMs) show promise in radiological diagnosis, but their performance may be affected by the context of the cases presented.
背景大语言模型（LLM）在放射学诊断中大有可为，但其性能可能会受到病例背景的影响。
{"title":"Influence of Prior Probability Information on Large Language Model Performance in Radiological Diagnosis","authors":"Takahiro Fukushima, Ryo Kurokawa, Akifumi Hagiwara, Yuki Sonoda, Yusuke Asari, Mariko Kurokawa, Jun Kanzawa, Wataru Gonoi, Osamu Abe","doi":"10.1101/2024.08.27.24312693","DOIUrl":"https://doi.org/10.1101/2024.08.27.24312693","url":null,"abstract":"<strong>Background</strong> Large language models (LLMs) show promise in radiological diagnosis, but their performance may be affected by the context of the cases presented.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"26 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181572","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-28DOI: 10.1101/2024.08.28.24312720
Alen Musisi, Rebecca Nakatudde, Oliver Namuwonge, Deborah Babirye, Ismail Kintu, Francis Olweny, Richard Malumba, Victoria Nakalanzi, Aloysius Gonzaga Mubuuke
Introduction/background: The heart is vital, and even minor dysfunctions can significantly impact the body. Cardiologists need always to determine heart size, which varies with physiological changes. Advanced measurement techniques are costly and often inaccessible to a common man. Measuring the cardiothoracic ratio (CTR) via conventional X-ray is a common and more affordable option, but there's a need for even cheaper alternatives Objective: To determine relationship between CTR and presenting clinical indications and to relate CTR to the body parameters to find an appropriate relationship that can be utilized in low resource facilities in determining heart size. Methodology: This cross-sectional study involved 386 patients undergoing chest radiographs at Mulago National Specialized Hospital's radiology department. Data were summarized using frequencies and percentages. Associations between the cardiothoracic ratio (CTR) and independent variables were analyzed using Pearson’s chi-square, Fisher’s exact test, Spearman’s correlation coefficient, simple linear regression, and multivariate regression. Statistical significance was set at a p-value of < 0.05. Results: The median cardiothoracic ratio (CTR) was 0.46, with an interquartile range of 0.42 to 0.50. Female patients had a higher CTR than males. Significant positive correlations were found between CTR; and BMI (p < 0.001, correlation 0.21), and BSA (p = 0.016, correlation 0.12), and BSI (p < 0.001, correlation 0.19). The diagnostic accuracy of a linear regression equation containing BSA as an estimator of CTR showed relatively fair performance compared to the linear regression equations with BSI and BMI. It showed sensitivity, specificity, and positive and negative predictive values of 29.2%, 86.0%, 63.6%, and 59.0% for males, and 8.3%, 98.1%, 75.0%, and 60.7% for females, respectively. Conclusion: BSA shows a moderately good relationship with CTR, while the influence of body habitus on CTR is minimal. Thus, using body parameters to predict CTR should be approached cautiously. We recommend conducting a similar study on a more diverse general population
{"title":"Cardiothoracic Ratio (CTR) Among Patients Presenting for Chest X-ray in Radiology Department at Mulago National Referral Hospital: A Patients’ Health Indicator for Clinical Application.","authors":"Alen Musisi, Rebecca Nakatudde, Oliver Namuwonge, Deborah Babirye, Ismail Kintu, Francis Olweny, Richard Malumba, Victoria Nakalanzi, Aloysius Gonzaga Mubuuke","doi":"10.1101/2024.08.28.24312720","DOIUrl":"https://doi.org/10.1101/2024.08.28.24312720","url":null,"abstract":"Introduction/background: The heart is vital, and even minor dysfunctions can significantly impact the body. Cardiologists need always to determine heart size, which varies with physiological changes. Advanced measurement techniques are costly and often inaccessible to a common man. Measuring the cardiothoracic ratio (CTR) via conventional X-ray is a common and more affordable option, but there's a need for even cheaper alternatives\u0000Objective: To determine relationship between CTR and presenting clinical indications and to relate CTR to the body parameters to find an appropriate relationship that can be utilized in low resource facilities in determining heart size.\u0000Methodology: This cross-sectional study involved 386 patients undergoing chest radiographs at Mulago National Specialized Hospital's radiology department. Data were summarized using frequencies and percentages. Associations between the cardiothoracic ratio (CTR) and independent variables were analyzed using Pearson’s chi-square, Fisher’s exact test, Spearman’s correlation coefficient, simple linear regression, and multivariate regression. Statistical significance was set at a p-value of < 0.05.\u0000Results: The median cardiothoracic ratio (CTR) was 0.46, with an interquartile range of 0.42 to 0.50. Female patients had a higher CTR than males. Significant positive correlations were found between CTR; and BMI (p < 0.001, correlation 0.21), and BSA (p = 0.016, correlation 0.12), and BSI (p < 0.001, correlation 0.19). The diagnostic accuracy of a linear regression equation containing BSA as an estimator of CTR showed relatively fair performance compared to the linear regression equations with BSI and BMI. It showed sensitivity, specificity, and positive and negative predictive values of 29.2%, 86.0%, 63.6%, and 59.0% for males, and 8.3%, 98.1%, 75.0%, and 60.7% for females, respectively.\u0000Conclusion: BSA shows a moderately good relationship with CTR, while the influence of body habitus on CTR is minimal. Thus, using body parameters to predict CTR should be approached cautiously. We recommend conducting a similar study on a more diverse general population","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"185 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142227565","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-28DOI: 10.1101/2024.08.27.24312653
Jawed Nawabi, Georg Lukas Baumgaertner, Sophia Schulze-Weddige, Andrea Dell'Orco, Andrea Morotti, Federico Mazzacane, Helge C Kniep, Frieder Schlunk, Maik FH Boehmer, Burakhan Akkurt, Tobias Orth, Jana-Sofie Weissflog, Maik Schumann, Peter Sporns, Michael Scheel, Uta Hanning, Jens Fiehler, Tobias Penzkofer
Purpose: To evaluate a nnU-Net-based deep learning for automated segmentation of intracerebral hemorrhage (ICH), intraventricular hemorrhage (IVH), and perihematomal edema (PHE) on noncontrast CT scans. Materials and Methods: Retrospective data from acute ICH patients admitted at four European stroke centers (2017-2019), along healthy controls (2022-2023), were analyzed. nnU-Net was trained (n=775) using a 5-fold cross-valiadtion approach, tested (n=189), and seperatly validated on internal (n=121), external (n=169), and diverse ICH etiologies (n=175) datasets. Interrater-validated ground truth served as the reference standard. Lesion detection, segmentation, and volumetric accuracy were measured, alongside time efficiency versus manual segmentation. Results: Test set results revealed high nnU-Net accuracy (median Dice Similartiy Coefficient (DSC): ICH 0.91, IVH 0.76, PHE 0.71) and volumetric correlation (ICH, IVH: r=0.99; PHE: r=0.92). Sensitivities were high (ICH, PHE: 99%; IVH: 97%), with IVH detection specificities and sensitivities >90% for volumes up to 0.2 ml. Anatomical-specific metrics showed higher performance for lobar and deep hemorrhages (median DSC 0.90 and 0.92, respectively) and lower for brainstem (median DSC 0.70). Concurrent hemorrhages did not affect accuracy, p> 0.05. Across validation sets, segmentation precision was consistent, especially for ICH (median DSC 0.85-0.90), with PHE slightly lower (median DSC 0.61-0.66) and IVH best in the second and third set (median DSC 0.80). Average processing time was 18.2 seconds versus 18.01 minutes manually. Conclusion: The nnU-Net provides reliable, time-efficient ICH, IVH, and PHE segmentation, validated across various clinical settings, with excellent anatomical-specific performance for lobar and deep hemorrhages. It shows promise for enhancing clinical workflow and research initiatives.
{"title":"Cross-Institutional European Evaluation and Validation of Automated Multilabel Segmentation for Acute Intracerebral Hemorrhage and Complications","authors":"Jawed Nawabi, Georg Lukas Baumgaertner, Sophia Schulze-Weddige, Andrea Dell'Orco, Andrea Morotti, Federico Mazzacane, Helge C Kniep, Frieder Schlunk, Maik FH Boehmer, Burakhan Akkurt, Tobias Orth, Jana-Sofie Weissflog, Maik Schumann, Peter Sporns, Michael Scheel, Uta Hanning, Jens Fiehler, Tobias Penzkofer","doi":"10.1101/2024.08.27.24312653","DOIUrl":"https://doi.org/10.1101/2024.08.27.24312653","url":null,"abstract":"Purpose: To evaluate a nnU-Net-based deep learning for automated segmentation of intracerebral hemorrhage (ICH), intraventricular hemorrhage (IVH), and perihematomal edema (PHE) on noncontrast CT scans. Materials and Methods: Retrospective data from acute ICH patients admitted at four European stroke centers (2017-2019), along healthy controls (2022-2023), were analyzed. nnU-Net was trained (n=775) using a 5-fold cross-valiadtion approach, tested (n=189), and seperatly validated on internal (n=121), external (n=169), and diverse ICH etiologies (n=175) datasets. Interrater-validated ground truth served as the reference standard. Lesion detection, segmentation, and volumetric accuracy were measured, alongside time efficiency versus manual segmentation. Results: Test set results revealed high nnU-Net accuracy (median Dice Similartiy Coefficient (DSC): ICH 0.91, IVH 0.76, PHE 0.71) and volumetric correlation (ICH, IVH: r=0.99; PHE: r=0.92). Sensitivities were high (ICH, PHE: 99%; IVH: 97%), with IVH detection specificities and sensitivities >90% for volumes up to 0.2 ml. Anatomical-specific metrics showed higher performance for lobar and deep hemorrhages (median DSC 0.90 and 0.92, respectively) and lower for brainstem (median DSC 0.70). Concurrent hemorrhages did not affect accuracy, p> 0.05. Across validation sets, segmentation precision was consistent, especially for ICH (median DSC 0.85-0.90), with PHE slightly lower (median DSC 0.61-0.66) and IVH best in the second and third set (median DSC 0.80). Average processing time was 18.2 seconds versus 18.01 minutes manually. Conclusion: The nnU-Net provides reliable, time-efficient ICH, IVH, and PHE segmentation, validated across various clinical settings, with excellent anatomical-specific performance for lobar and deep hemorrhages. It shows promise for enhancing clinical workflow and research initiatives.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"9 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181570","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-28DOI: 10.1101/2024.08.27.24312482
Yazdan Salimi, Zahra Mansouri, Isaac Shiri, Ismini Mainta, Habib Zaidi
Introduction: The common approach for organ segmentation in hybrid imaging relies on co-registered CT (CTAC) images. This method, however, presents several limitations in real clinical workflows where mismatch between PET and CT images are very common. Moreover, low-dose CTAC images have poor quality, thus challenging the segmentation task. Recent advances in CT-less PET imaging further highlight the necessity for an effective PET organ segmentation pipeline that does not rely on CT images. Therefore, the goal of this study was to develop a CT-less multi-tracer PET segmentation framework. Methods: We collected 2062 PET/CT images from multiple scanners. The patients were injected with either 18F-FDG (1487) or 68Ga-PSMA (575). PET/CT images with any kind of mismatch between PET and CT images were detected through visual assessment and excluded from our study. Multiple organs were delineated on CT components using previously trained in-house developed nnU-Net models. The segmentation masks were resampled to co-registered PET images and used to train four different deep-learning models using different images as input, including non-corrected PET (PET-NC) and attenuation and scatter-corrected PET (PET-ASC) for 18F-FDG (tasks #1 and #2, respectively using 22 organs) and PET-NC and PET-ASC for 68Ga tracers (tasks #3 and #4, respectively, using 15 organs). The models performance was evaluated in terms of Dice coefficient, Jaccard index, and segment volume difference. Results: The average Dice coefficient over all organs was 0.81±0.15, 0.82±0.14, 0.77±0.17, and 0.79±0.16 for tasks #1, #2, #3, and #4, respectively. PET-ASC models outperformed PET-NC models (P-value < 0.05). The highest Dice values were achieved for the brain (0.93 to 0.96 in all four tasks), whereas the lowest values were achieved for small organs, such as the adrenal glands. The trained models showed robust performance on dynamic noisy images as well. Conclusion: Deep learning models allow high performance multi-organ segmentation for two popular PET tracers without the use of CT information. These models may tackle the limitations of using CT segmentation in PET/CT image quantification, kinetic modeling, radiomics analysis, dosimetry, or any other tasks that require organ segmentation masks.
{"title":"Deep Learning-powered CT-less Multi-tracer Organ Segmentation from PET Images: A solution for unreliable CT segmentation in PET/CT Imaging","authors":"Yazdan Salimi, Zahra Mansouri, Isaac Shiri, Ismini Mainta, Habib Zaidi","doi":"10.1101/2024.08.27.24312482","DOIUrl":"https://doi.org/10.1101/2024.08.27.24312482","url":null,"abstract":"Introduction: The common approach for organ segmentation in hybrid imaging relies on co-registered CT (CTAC) images. This method, however, presents several limitations in real clinical workflows where mismatch between PET and CT images are very common. Moreover, low-dose CTAC images have poor quality, thus challenging the segmentation task. Recent advances in CT-less PET imaging further highlight the necessity for an effective PET organ segmentation pipeline that does not rely on CT images. Therefore, the goal of this study was to develop a CT-less multi-tracer PET segmentation framework.\u0000Methods: We collected 2062 PET/CT images from multiple scanners. The patients were injected with either 18F-FDG (1487) or 68Ga-PSMA (575). PET/CT images with any kind of mismatch between PET and CT images were detected through visual assessment and excluded from our study. Multiple organs were delineated on CT components using previously trained in-house developed nnU-Net models. The segmentation masks were resampled to co-registered PET images and used to train four different deep-learning models using different images as input, including non-corrected PET (PET-NC) and attenuation and scatter-corrected PET (PET-ASC) for 18F-FDG (tasks #1 and #2, respectively using 22 organs) and PET-NC and PET-ASC for 68Ga tracers (tasks #3 and #4, respectively, using 15 organs). The models performance was evaluated in terms of Dice coefficient, Jaccard index, and segment volume difference.\u0000Results: The average Dice coefficient over all organs was 0.81±0.15, 0.82±0.14, 0.77±0.17, and 0.79±0.16 for tasks #1, #2, #3, and #4, respectively. PET-ASC models outperformed PET-NC models (P-value < 0.05). The highest Dice values were achieved for the brain (0.93 to 0.96 in all four tasks), whereas the lowest values were achieved for small organs, such as the adrenal glands. The trained models showed robust performance on dynamic noisy images as well.\u0000Conclusion: Deep learning models allow high performance multi-organ segmentation for two popular PET tracers without the use of CT information. These models may tackle the limitations of using CT segmentation in PET/CT image quantification, kinetic modeling, radiomics analysis, dosimetry, or any other tasks that require organ segmentation masks.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"17 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181573","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-23DOI: 10.1101/2024.08.19.24312117
Katy Vecchiato, Chiara Casella, Ayse Sila Dokumaci, Olivia Carney, Jon O. Cleary, Pierluigi Di Ciò, Michela Cleri, Kathleen Colford, Rory J. Piper, Tomoki Arichi Arichi, Michael Eyre, Fraser Aitken, Raphael Tomi-Tricot, Tom Wilkinson, Colm J. McGinnity, Sharon L. Giles, Shaihan Malik, Alexander Hammers, Philippa Bridgen, David W Carmichael, Jonathan O'Muircheartaigh
Background and Objectives: Epileptogenic lesions in focal epilepsy can be subtle or undetected on conventional brain MRI. Ultra-high field (7T) MRI offers higher spatial resolution, contrast and signal-to-noise ratio compared to conventional imaging systems and has shown promise in the pre-surgical evaluation of adult focal epilepsy. However, the utility of ultra-high field MRI in paediatric focal epilepsy, where malformations of cortical development are more common, is unclear. This study compared 7T to conventional 3T MRI in children with epilepsy by comparing: (i) scan tolerability; (ii) radiological image quality; (iii) lesion yield. Materials and Methods: Children with drug-resistant focal epilepsy and healthy controls were recruited prospectively and imaged at both 3T and 7T. Safety and tolerability during scanning was assessed via a questionnaire. Image quality was evaluated by an expert paediatric neuroradiologist and estimated quantitatively by comparing cortical thickness between field strengths. To assess lesion detection yield of 7T MRI, a multi-disciplinary team jointly reviewed patients' images. Results: 41 patients (8-17 years, mean=12.6 years, 22 male) and 22 healthy controls (8-17 years, mean=11.7 years, 15 male) were recruited. All children completed the scan, with no significant adverse events. Higher discomfort due to dizziness was reported at 7T (p=0.02), with side-effects more frequently noted in younger children (p=0.02). However, both field strengths were generally well-tolerated and side-effects were transient. 7T images had increased inhomogeneity and artefacts compared to those obtained at 3T. Cortical thickness measurements were significantly thinner at 7T (p<0.001). 8/26 (31%) patients had new lesions identified at 7T which were not identified at 3T, influencing the surgical management in 4/26 (15%). Discussion: 7T MRI in children with epilepsy is feasible, well-tolerated and is associated with a 31% improvement in lesion detection rates.
{"title":"High-Field 7T MRI in a drug-resistant paediatric epilepsy cohort: image comparison and radiological outcomes","authors":"Katy Vecchiato, Chiara Casella, Ayse Sila Dokumaci, Olivia Carney, Jon O. Cleary, Pierluigi Di Ciò, Michela Cleri, Kathleen Colford, Rory J. Piper, Tomoki Arichi Arichi, Michael Eyre, Fraser Aitken, Raphael Tomi-Tricot, Tom Wilkinson, Colm J. McGinnity, Sharon L. Giles, Shaihan Malik, Alexander Hammers, Philippa Bridgen, David W Carmichael, Jonathan O'Muircheartaigh","doi":"10.1101/2024.08.19.24312117","DOIUrl":"https://doi.org/10.1101/2024.08.19.24312117","url":null,"abstract":"Background and Objectives: Epileptogenic lesions in focal epilepsy can be subtle or undetected on conventional brain MRI. Ultra-high field (7T) MRI offers higher spatial resolution, contrast and signal-to-noise ratio compared to conventional imaging systems and has shown promise in the pre-surgical evaluation of adult focal epilepsy. However, the utility of ultra-high field MRI in paediatric focal epilepsy, where malformations of cortical development are more common, is unclear. This study compared 7T to conventional 3T MRI in children with epilepsy by comparing: (i) scan tolerability; (ii) radiological image quality; (iii) lesion yield. Materials and Methods: Children with drug-resistant focal epilepsy and healthy controls were recruited prospectively and imaged at both 3T and 7T. Safety and tolerability during scanning was assessed via a questionnaire. Image quality was evaluated by an expert paediatric neuroradiologist and estimated quantitatively by comparing cortical thickness between field strengths. To assess lesion detection yield of 7T MRI, a multi-disciplinary team jointly reviewed patients' images. Results: 41 patients (8-17 years, mean=12.6 years, 22 male) and 22 healthy controls (8-17 years, mean=11.7 years, 15 male) were recruited. All children completed the scan, with no significant adverse events. Higher discomfort due to dizziness was reported at 7T (p=0.02), with side-effects more frequently noted in younger children (p=0.02). However, both field strengths were generally well-tolerated and side-effects were transient. 7T images had increased inhomogeneity and artefacts compared to those obtained at 3T. Cortical thickness measurements were significantly thinner at 7T (p<0.001). 8/26 (31%) patients had new lesions identified at 7T which were not identified at 3T, influencing the surgical management in 4/26 (15%). Discussion: 7T MRI in children with epilepsy is feasible, well-tolerated and is associated with a 31% improvement in lesion detection rates.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"87 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181588","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-23DOI: 10.1101/2024.08.23.24312461
Li Zhang, Basu Jindal, Ahmed Alaa, Robert Weinreb, David Wilson, Eran Segal, James Zou, Pengtao Xie
Semantic segmentation of medical images is pivotal in applications like disease diagnosis and treatment planning. While deep learning has excelled in automating this task, a major hurdle is the need for numerous annotated segmentation masks, which are resource-intensive to produce due to the required expertise and time. This scenario often leads to ultra low-data regimes, where annotated images are extremely limited, posing significant challenges for the generalization of conventional deep learning methods on test images. To address this, we introduce a generative deep learning framework, which uniquely generates high-quality paired segmentation masks and medical images, serving as auxiliary data for training robust models in data-scarce environments. Unlike traditional generative models that treat data generation and segmentation model training as separate processes, our method employs multi-level optimization for end-to-end data generation. This approach allows segmentation performance to directly influence the data generation process, ensuring that the generated data is specifically tailored to enhance the performance of the segmentation model. Our method demonstrated strong generalization performance across 9 diverse medical image segmentation tasks and on 16 datasets, in ultra-low data regimes, spanning various diseases, organs, and imaging modalities. When applied to various segmentation models, it achieved performance improvements of 10-20% (absolute), in both same-domain and out-of-domain scenarios. Notably, it requires 8 to 20 times less training data than existing methods to achieve comparable results. This advancement significantly improves the feasibility and cost-effectiveness of applying deep learning in medical imaging, particularly in scenarios with limited data availability.
{"title":"Generative AI Enables Medical Image Segmentation in Ultra Low-Data Regimes","authors":"Li Zhang, Basu Jindal, Ahmed Alaa, Robert Weinreb, David Wilson, Eran Segal, James Zou, Pengtao Xie","doi":"10.1101/2024.08.23.24312461","DOIUrl":"https://doi.org/10.1101/2024.08.23.24312461","url":null,"abstract":"Semantic segmentation of medical images is pivotal in applications like disease diagnosis and treatment planning. While deep learning has excelled in automating this task, a major hurdle is the need for numerous annotated segmentation masks, which are resource-intensive to produce due to the required expertise and time. This scenario often leads to ultra low-data regimes, where annotated images are extremely limited, posing significant challenges for the generalization of conventional deep learning methods on test images. To address this, we introduce a generative deep learning framework, which uniquely generates high-quality paired segmentation masks and medical images, serving as auxiliary data for training robust models in data-scarce environments. Unlike traditional generative models that treat data generation and segmentation model training as separate processes, our method employs multi-level optimization for end-to-end data generation. This approach allows segmentation performance to directly influence the data generation process, ensuring that the generated data is specifically tailored to enhance the performance of the segmentation model. Our method demonstrated strong generalization performance across 9 diverse medical image segmentation tasks and on 16 datasets, in ultra-low data regimes, spanning various diseases, organs, and imaging modalities. When applied to various segmentation models, it achieved performance improvements of 10-20% (absolute), in both same-domain and out-of-domain scenarios. Notably, it requires 8 to 20 times less training data than existing methods to achieve comparable results. This advancement significantly improves the feasibility and cost-effectiveness of applying deep learning in medical imaging, particularly in scenarios with limited data availability.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"18 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181587","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-22DOI: 10.1101/2024.08.16.24312127
Harriet Nalubega Kisembo, Richard Malumba, Ezra Kato Nsereko, Deborah Babirye, Victoria Nakalanzi, Francis Xavier Kasujja, Elsie Kiguli Malwadde, Elizeus Rutebemberwa, Simon Kasasa, Dina Husseiny Salama, Michael Grace Kawooya
Background Multi-Detector Computed Tomography (MDCT) has revolutionized healthcare delivery, significantly improving diagnostic accuracy and patient outcomes in various clinical settings. However, the overuse of CT examinations (CTEs), especially in resource-limited settings (RLS), poses a substantial public health challenge. Inappropriately performed CTEs, particularly among children and young adults, expose these vulnerable populations to unnecessary radiation risks, with 20%-50% of CTEs deemed inappropriate, and 10%-20% involving children. Despite the existence of evidence-based interventions like clinical imaging guidelines (CIGs) to curb this overuse, their availability and effectiveness in RLS are not well established. Objective This study aimed to determine the impact of continuous medical education (CME) and the introduction of clinical imaging guidelines (CIGs) on the appropriateness of CT utilization among children and young adults in selected hospitals in Uganda. Materials and Methods A before-and-after study design was employed to assess the effect of an intervention comprising CME and CIGs on appropriate CTE utilization. The intervention targeted healthcare providers (HCPs) across six public and private tertiary hospitals with available CT services over a 12-month period. Baseline data indicated a high prevalence of inappropriate CTEs among the target population. The proportion of CTEs performed for various body regions (head, paranasal sinuses, chest, abdomen, spine, trauma) and their appropriateness were retrospectively analyzed before and after the intervention, using the European Society of Radiology's iGuide and pre-intervention study results as benchmarks. Results Post-intervention, the total number of CTEs performed increased by 33% (909 vs. 1210), with a 30% increase in public hospitals (300 vs. 608, p < 0.001) and a 41% increase in private-for-profit hospitals (91 vs. 238, p = 0.037). Specific increases were observed in head CTs (19%, 746 vs. 890, p < 0.0001) and contrasted studies (252%, 113 vs. 410, p < 0.0001). Conversely, CTEs for trauma decreased by 8% (499 vs. 458, p < 0.0001). Despite these changes, the overall proportion of inappropriate CTEs increased by 15% (38% vs. 44%, p < 0.001), with a 28% increase in inappropriate contrasted examinations (25% vs. 53%, p < 0.001) and a 13% increase in non-trauma cases (66% vs. 79%, p < 0.001). Notably, inappropriate CTEs for non-contrasted and trauma-related cases reduced by 28% (75% vs. 47%, p < 0.001) and 31% (34% vs. 14%, p = 0.0001), respectively. Conclusion The findings underscore the potential of CME and the adaptation of CIGs from high-resource settings to enhance the appropriateness of CT utilization in RLS. While the intervention notably reduced inappropriate trauma-related and non-contrasted CTEs, it also highlighted the complexity of achieving consistent improvements across all examination types. Further research is recommended to explore the
{"title":"Effect of continuous medical education and clinical imaging guidelines on reducing inappropriate computerized tomography utilization among children and young patients in a resource -limited settings: A before-and-after study","authors":"Harriet Nalubega Kisembo, Richard Malumba, Ezra Kato Nsereko, Deborah Babirye, Victoria Nakalanzi, Francis Xavier Kasujja, Elsie Kiguli Malwadde, Elizeus Rutebemberwa, Simon Kasasa, Dina Husseiny Salama, Michael Grace Kawooya","doi":"10.1101/2024.08.16.24312127","DOIUrl":"https://doi.org/10.1101/2024.08.16.24312127","url":null,"abstract":"Background Multi-Detector Computed Tomography (MDCT) has revolutionized healthcare delivery, significantly improving diagnostic accuracy and patient outcomes in various clinical settings. However, the overuse of CT examinations (CTEs), especially in resource-limited settings (RLS), poses a substantial public health challenge. Inappropriately performed CTEs, particularly among children and young adults, expose these vulnerable populations to unnecessary radiation risks, with 20%-50% of CTEs deemed inappropriate, and 10%-20% involving children. Despite the existence of evidence-based interventions like clinical imaging guidelines (CIGs) to curb this overuse, their availability and effectiveness in RLS are not well established.\u0000Objective\u0000This study aimed to determine the impact of continuous medical education (CME) and the introduction of clinical imaging guidelines (CIGs) on the appropriateness of CT utilization among children and young adults in selected hospitals in Uganda.\u0000Materials and Methods\u0000A before-and-after study design was employed to assess the effect of an intervention comprising CME and CIGs on appropriate CTE utilization. The intervention targeted healthcare providers (HCPs) across six public and private tertiary hospitals with available CT services over a 12-month period. Baseline data indicated a high prevalence of inappropriate CTEs among the target population. The proportion of CTEs performed for various body regions (head, paranasal sinuses, chest, abdomen, spine, trauma) and their appropriateness were retrospectively analyzed before and after the intervention, using the European Society of Radiology's iGuide and pre-intervention study results as benchmarks.\u0000Results Post-intervention, the total number of CTEs performed increased by 33% (909 vs. 1210), with a 30% increase in public hospitals (300 vs. 608, p < 0.001) and a 41% increase in private-for-profit hospitals (91 vs. 238, p = 0.037). Specific increases were observed in head CTs (19%, 746 vs. 890, p < 0.0001) and contrasted studies (252%, 113 vs. 410, p < 0.0001). Conversely, CTEs for trauma decreased by 8% (499 vs. 458, p < 0.0001). Despite these changes, the overall proportion of inappropriate CTEs increased by 15% (38% vs. 44%, p < 0.001), with a 28% increase in inappropriate contrasted examinations (25% vs. 53%, p < 0.001) and a 13% increase in non-trauma cases (66% vs. 79%, p < 0.001). Notably, inappropriate CTEs for non-contrasted and trauma-related cases reduced by 28% (75% vs. 47%, p < 0.001) and 31% (34% vs. 14%, p = 0.0001), respectively.\u0000Conclusion\u0000The findings underscore the potential of CME and the adaptation of CIGs from high-resource settings to enhance the appropriateness of CT utilization in RLS. While the intervention notably reduced inappropriate trauma-related and non-contrasted CTEs, it also highlighted the complexity of achieving consistent improvements across all examination types. Further research is recommended to explore the ","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"35 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181589","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-22DOI: 10.1101/2024.08.22.24312200
Jamal Esmaelpoor, Tommy Peng, Beth Jelfs, Darren Mao, Maureen J. Shader, Colette M. McKay
Objective: Despite evidence that cross-modal effects after hearing loss and cochlear implantation are primarily conveyed through synaptic gain and efficacy rather than reorganized fiber tracts, few studies have assessed cross-modal functional connectivity (CMFC) to evaluate plasticity. This study, inspired by the psychophysiological interactions (PPI) method, addresses its limitations and provides a robust approach to evaluating task-induced CMFC. Design: Twenty-two post-lingually deafened, newly implanted adult cochlear implant (CI) recipients with severe hearing loss in the contralateral ear and 17 normal-hearing (NH) subjects participated. The experiment included audio-only and visual-only speech tasks, with resting-state FC as a baseline. Functional near-infrared spectroscopy (fNIRS) measured brain imaging data one month and one year post-implantation. CI users' speech understanding performance was evaluated one year after implantation. Results: A negative correlation was found between average contralateral task-induced CMFC and speech outcomes, particularly in links from the angular gyrus (AG), both one month and one year post-activation. Plastic changes showed higher task-induced CMFC in AG compared to the superior temporal gyrus (STG), aligning with neural efficiency principles. Task-induced CMFC remained elevated in CI users compared to NH cohorts even after one year. Conclusion: Task-induced CMFC can serve as a significant marker of cross-modal plasticity and speech performance in CI recipients, indicating increased reliance on cross-modal processing in one year after implantation.
目的：尽管有证据表明，听力损失和人工耳蜗植入后的跨模态效应主要是通过突触增益和功效而非重组纤维束传递的，但很少有研究通过评估跨模态功能连接（CMFC）来评估可塑性。本研究受心理生理学相互作用（PPI）方法的启发，解决了该方法的局限性，为评估任务诱导的CMFC提供了一种可靠的方法：22名新近植入人工耳蜗（CI）且对侧耳听力严重受损的语后聋成人受试者和17名听力正常（NH）的受试者参加了实验。实验包括纯音频和纯视觉言语任务，以静息态 FC 为基线。功能性近红外光谱（fNIRS）测量了植入后一个月和一年的脑成像数据。植入 CI 一年后，对使用者的语音理解能力进行了评估：结果：在植入一个月和一年后，发现对侧任务诱导的 CMFC 平均值与语音结果之间存在负相关，尤其是在角回（AG）的链接中。塑性变化显示，与颞上回（STG）相比，AG 的任务诱导 CMFC 更高，这符合神经效率原则。即使在一年后，CI使用者的任务诱导CMFC仍然高于NH组群：任务诱导的 CMFC 可作为 CI 接受者跨模态可塑性和语言表达能力的重要标志，表明植入一年后对跨模态处理的依赖性增加。
{"title":"Cross-modal Functional Plasticity after Cochlear-implantation","authors":"Jamal Esmaelpoor, Tommy Peng, Beth Jelfs, Darren Mao, Maureen J. Shader, Colette M. McKay","doi":"10.1101/2024.08.22.24312200","DOIUrl":"https://doi.org/10.1101/2024.08.22.24312200","url":null,"abstract":"Objective: Despite evidence that cross-modal effects after hearing loss and cochlear implantation are primarily conveyed through synaptic gain and efficacy rather than reorganized fiber tracts, few studies have assessed cross-modal functional connectivity (CMFC) to evaluate plasticity. This study, inspired by the psychophysiological interactions (PPI) method, addresses its limitations and provides a robust approach to evaluating task-induced CMFC.\u0000Design: Twenty-two post-lingually deafened, newly implanted adult cochlear implant (CI) recipients with severe hearing loss in the contralateral ear and 17 normal-hearing (NH) subjects participated. The experiment included audio-only and visual-only speech tasks, with resting-state FC as a baseline. Functional near-infrared spectroscopy (fNIRS) measured brain imaging data one month and one year post-implantation. CI users' speech understanding performance was evaluated one year after implantation.\u0000Results: A negative correlation was found between average contralateral task-induced CMFC and speech outcomes, particularly in links from the angular gyrus (AG), both one month and one year post-activation. Plastic changes showed higher task-induced CMFC in AG compared to the superior temporal gyrus (STG), aligning with neural efficiency principles. Task-induced CMFC remained elevated in CI users compared to NH cohorts even after one year.\u0000Conclusion: Task-induced CMFC can serve as a significant marker of cross-modal plasticity and speech performance in CI recipients, indicating increased reliance on cross-modal processing in one year after implantation.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"54 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181590","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-22DOI: 10.1101/2024.08.22.24312425
Hoda Kalabizadeh, Ludovica Griffanti, Pak Hei Yeung, Natalie Voets, Grace Gillis, Clare E Mackay, Ana IL Namburete, Nicola K Dinsdale, Konstantinos Kamnitsas
Brain atrophy assessment in MRI, particularly of the hippocampus, is commonly used to support diagnosis and monitoring of dementia. Consequently, there is a demand for accurate automated hippocampus quantification. Most existing segmentation methods have been developed and validated on research datasets and, therefore, may not be appropriate for clinical MR images and populations, leading to potential gaps between dementia research and clinical practice. In this study, we investigated the performance of segmentation models trained on research data that were style-transferred to resemble clinical scans. Our results highlighted the importance of intensity normalisation methods in MRI segmentation, and their relation to domain shift and style-transfer. We found that whilst normalising intensity based on min and max values, commonly used in generative MR harmonisation methods, may create a need for style transfer, Z-score normalisation effectively maintains style consistency, and optimises performance. Moreover, we show for our datasets spatial augmentations are more beneficial than style harmonisation. Thus, emphasising robust normalisation techniques and spatial augmentation significantly improves MRI hippocampus segmentation.
核磁共振成像中的脑萎缩评估，尤其是海马体的萎缩评估，通常用于痴呆症的诊断和监测。因此，需要对海马体进行精确的自动量化。现有的大多数分割方法都是在研究数据集上开发和验证的，因此可能不适合临床磁共振图像和人群，导致痴呆症研究和临床实践之间可能存在差距。在本研究中，我们研究了在研究数据上训练的分割模型的性能，这些数据经过样式转换后与临床扫描数据相似。我们的研究结果强调了核磁共振成像分割中强度归一化方法的重要性，以及它们与领域转移和风格转换的关系。我们发现，虽然基于最小值和最大值的强度归一化（通常用于生成式磁共振协调方法）可能会产生风格转换需求，但 Z 分数归一化能有效保持风格一致性，并优化性能。此外，我们的数据集显示，空间增强比风格协调更有益。因此，强调稳健的归一化技术和空间增强技术能显著提高磁共振成像海马区块的分割效果。
{"title":"Is Your Style Transfer Doing Anything Useful? An Investigation Into Hippocampus Segmentation and the Role of Preprocessing","authors":"Hoda Kalabizadeh, Ludovica Griffanti, Pak Hei Yeung, Natalie Voets, Grace Gillis, Clare E Mackay, Ana IL Namburete, Nicola K Dinsdale, Konstantinos Kamnitsas","doi":"10.1101/2024.08.22.24312425","DOIUrl":"https://doi.org/10.1101/2024.08.22.24312425","url":null,"abstract":"Brain atrophy assessment in MRI, particularly of the hippocampus, is commonly used to support diagnosis and monitoring of dementia. Consequently, there is a demand for accurate automated hippocampus quantification. Most existing segmentation methods have been developed and validated on research datasets and, therefore, may not be appropriate for clinical MR images and populations, leading to potential gaps between dementia research and clinical practice. In this study, we investigated the performance of segmentation models trained on research data that were style-transferred to resemble clinical scans. Our results highlighted the importance of intensity normalisation methods in MRI segmentation, and their relation to domain shift and style-transfer. We found that whilst normalising intensity based on min and max values, commonly used in generative MR harmonisation methods, may create a need for style transfer, Z-score normalisation effectively maintains style consistency, and optimises performance. Moreover, we show for our datasets spatial augmentations are more beneficial than style harmonisation. Thus, emphasising robust normalisation techniques and spatial augmentation significantly improves MRI hippocampus segmentation.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"113 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142223836","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-08-21DOI: 10.1101/2024.08.21.24312346
Patricio A. Pincheira, Jong H. Kim, Paul W. Hodges
Objective This study aimed to develop a machine learning method for characterizing muscle composition on ultrasound imaging, focusing on pixel-level quantification of connective tissue using texture analysis. Methods Ultrasound images of the multifidus muscle from 20 healthy young adults were included in the analysis. Texture features including Local Binary Patterns, Histograms of Oriented Gradients, Grey Level Co-occurrence Matrix, and Discrete Wavelet Transforms, were extracted from the images across multiple scales. Within a positive-unlabeled machine learning framework, two competing models, Bagging Support Vector Machine and Random Forests with Recursive Greedy Risk Minimization were trained for each texture and scale. The outputs of the texture-based pixel-level classification were compared to traditional echo intensity-based methods. Metrics such as the F-measure were employed to evaluate the models' performance. Expert consensus was utilised to evaluate the accuracy of the classified images and identify the best-performing combination of model, texture, and scale. Results Expert evaluation identified the Bagging Support Vector Machine model trained with Local Binary Pattern histograms extracted at a scale of 9x9 pixel region of interest as the best combination for accurately classifying connective tissue-like pixels (F-measure= 0.88). The proposed method demonstrated high repeatability (intraclass correlation coefficient= 0.92) and robustness to echo intensity variations, outperforming traditional echo intensity-based methods. Conclusion This approach offers a valid method for pixel-level quantification of intramuscular connective tissue from ultrasound images. It overcomes the limitations of traditional analyses relying on echo intensity and demonstrates robustness against variations in echo intensity, representing an operator-independent advancement in ultrasound-based muscle composition analysis.
目的 本研究旨在开发一种机器学习方法来描述超声波成像上的肌肉成分，重点是利用纹理分析对结缔组织进行像素级量化。方法 分析对象包括 20 名健康年轻人的多裂肌超声图像。从图像中提取了多个尺度的纹理特征，包括局部二进制模式、定向梯度直方图、灰度共现矩阵和离散小波变换。在正向无标记机器学习框架内，针对每种纹理和尺度训练了两个竞争模型，即支持向量机（Bagging Support Vector Machine）和随机森林（Random Forests with Recursive Greedy Risk Minimization）。基于纹理的像素级分类输出结果与传统的基于回声强度的方法进行了比较。采用 F 测量等指标来评估模型的性能。专家共识用于评估分类图像的准确性，并确定模型、纹理和尺度的最佳组合。结果 专家评估认为，使用按 9x9 像素感兴趣区比例提取的局部二进制模式直方图训练的袋式支持向量机模型是准确分类结缔组织类像素的最佳组合（F-measure= 0.88）。所提出的方法重复性高（类内相关系数= 0.92），对回波强度变化的鲁棒性强，优于传统的基于回波强度的方法。结论 该方法为从超声图像中量化肌肉内结缔组织提供了一种有效的像素级方法。它克服了传统分析法依赖回声强度的局限性，对回声强度的变化表现出很强的鲁棒性，代表了超声肌肉成分分析中一种不受操作者影响的进步。
{"title":"Machine Learning-Based Pixel-Level Quantification of Intramuscular Connective Tissue using Ultrasound Texture Analysis","authors":"Patricio A. Pincheira, Jong H. Kim, Paul W. Hodges","doi":"10.1101/2024.08.21.24312346","DOIUrl":"https://doi.org/10.1101/2024.08.21.24312346","url":null,"abstract":"Objective This study aimed to develop a machine learning method for characterizing muscle composition on ultrasound imaging, focusing on pixel-level quantification of connective tissue using texture analysis. Methods Ultrasound images of the multifidus muscle from 20 healthy young adults were included in the analysis. Texture features including Local Binary Patterns, Histograms of Oriented Gradients, Grey Level Co-occurrence Matrix, and Discrete Wavelet Transforms, were extracted from the images across multiple scales. Within a positive-unlabeled machine learning framework, two competing models, Bagging Support Vector Machine and Random Forests with Recursive Greedy Risk Minimization were trained for each texture and scale. The outputs of the texture-based pixel-level classification were compared to traditional echo intensity-based methods. Metrics such as the F-measure were employed to evaluate the models' performance. Expert consensus was utilised to evaluate the accuracy of the classified images and identify the best-performing combination of model, texture, and scale. Results Expert evaluation identified the Bagging Support Vector Machine model trained with Local Binary Pattern histograms extracted at a scale of 9x9 pixel region of interest as the best combination for accurately classifying connective tissue-like pixels (F-measure= 0.88). The proposed method demonstrated high repeatability (intraclass correlation coefficient= 0.92) and robustness to echo intensity variations, outperforming traditional echo intensity-based methods. Conclusion This approach offers a valid method for pixel-level quantification of intramuscular connective tissue from ultrasound images. It overcomes the limitations of traditional analyses relying on echo intensity and demonstrates robustness against variations in echo intensity, representing an operator-independent advancement in ultrasound-based muscle composition analysis.","PeriodicalId":501358,"journal":{"name":"medRxiv - Radiology and Imaging","volume":"117 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142181592","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：