Radiography (London, England : 1995) · 2025
The impact of intensity discretisation and filtration on the performance of the radiomic and machine learning models in brain metastasis patients treated with gamma knife radiosurgery.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Umaru A, Manan H A, Kumar R K A, Hamsan S K, Yahya N
Аннотация
INTRODUCTION: Image preprocessing is crucial for optimizing radiomics feature extraction, however, inconsistencies in the implementation process and a lack of universally accepted methods lead to diverse approaches. This study evaluates the impact of radiomics and machine learning (ML) performance in brain metastasis.
METHODS: In this retrospective study, 108 lesions (regions of interest, ROI) were analysed. Using contrast-enhanced T1-weighted (T1W) images resampled to 1 × 1x1mm3, radiomics features were extracted using various fixed bin sizes (8, 16, 32, 64, 128, 256) and bin numbers (1, 5, 10, 25, 50) with relative ROI (max-min) intensity rescaling. Intensity rescaling methods, including 64 bins with relative mean ROI ±3SD and relative min-max rescaling, were applied. Additionally, four filters (Laplacian of Gaussian (LOG), Wavelet, Laws, and Mean) were tested. Feature selection was conducted using correlation analysis and Least Absolute Shrinkage and Selection Operator (LASSO) regression, yielding 11-18 features. K-nearest neighbors (KNN) was used for classification, and Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance.
RESULTS: The 32-bin model achieved the highest accuracy (70 %) and AUC (0.70; 95 % CI: 0.51-0.87) among the bin sizes, while the 10-bin model performed best among bin numbers, with accuracy of 79 % and AUC of 0.69 (95 % CI: 0.45-0.89). No significant difference was found in accuracy (p = 0.278) and AUC (p = 0.288) between bin sizes and numbers. Mean relative ROI ±3SD rescaling improved accuracy (73 % vs 61 %) and AUC (0.74 vs 0.60). The LOG and mean filters demonstrated superior performance.
CONCLUSION: Image preprocessing choices significantly influence radiomics features and ML performance. Standardized preprocessing is essential for enhancing reproducibility.
IMPLICATIONS FOR PRACTICE: Standardizing preprocessing methods can improve the reliability and generalizability of radiomics models, with potential applications in clinical decision-making for brain metastasis treatment.
METHODS: In this retrospective study, 108 lesions (regions of interest, ROI) were analysed. Using contrast-enhanced T1-weighted (T1W) images resampled to 1 × 1x1mm3, radiomics features were extracted using various fixed bin sizes (8, 16, 32, 64, 128, 256) and bin numbers (1, 5, 10, 25, 50) with relative ROI (max-min) intensity rescaling. Intensity rescaling methods, including 64 bins with relative mean ROI ±3SD and relative min-max rescaling, were applied. Additionally, four filters (Laplacian of Gaussian (LOG), Wavelet, Laws, and Mean) were tested. Feature selection was conducted using correlation analysis and Least Absolute Shrinkage and Selection Operator (LASSO) regression, yielding 11-18 features. K-nearest neighbors (KNN) was used for classification, and Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance.
RESULTS: The 32-bin model achieved the highest accuracy (70 %) and AUC (0.70; 95 % CI: 0.51-0.87) among the bin sizes, while the 10-bin model performed best among bin numbers, with accuracy of 79 % and AUC of 0.69 (95 % CI: 0.45-0.89). No significant difference was found in accuracy (p = 0.278) and AUC (p = 0.288) between bin sizes and numbers. Mean relative ROI ±3SD rescaling improved accuracy (73 % vs 61 %) and AUC (0.74 vs 0.60). The LOG and mean filters demonstrated superior performance.
CONCLUSION: Image preprocessing choices significantly influence radiomics features and ML performance. Standardized preprocessing is essential for enhancing reproducibility.
IMPLICATIONS FOR PRACTICE: Standardizing preprocessing methods can improve the reliability and generalizability of radiomics models, with potential applications in clinical decision-making for brain metastasis treatment.