Related Experiment Video
Updated: Jul 11, 2025
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Differentiation of Sinonasal NKT From Diffuse Large B-Cell Lymphoma Using Machine Learning and MRI-Based Radiomics
Yiyin Zhang1, Naier Lin1, Hanyu Xiao1
1From the Department of Radiology, Eye and ENT Hospital of Fudan University.
Purpose:
The aim of this study was to construct and validate a noninvasive radiomics method based on magnetic resonance imaging to differentiate sinonasal extranodal natural killer/T-cell lymphoma from diffuse large B-cell lymphoma.
Methods:
We collected magnetic resonance imaging scans, including contrast-enhanced T1-weighted imaging and T2-weighted imaging, from 133 patients with non-Hodgkin lymphoma (103 sinonasal extranodal natural killer/T-cell lymphoma and 30 diffuse large B-cell lymphoma) and randomly split them into training and testing cohorts at a ratio of 7:3. Clinical characteristics and image performance were analyzed to build a logistic regression clinical-image model. The radiomics features were extracted on contrast-enhanced T1-weighted imaging and T2-weighted imaging images. Maximum relevance minimum redundancy, selectKbest, and the least absolute shrinkage and selection operator algorithms (LASSO) were applied for feature selection after balancing the training set. Five machine learning classifiers were used to construct the single and combined sequences radiomics models. Sensitivity, specificity, accuracy, precision, F1score, the area under receiver operating characteristic curve, and the area under precision-recall curve were compared between the 15 models and the clinical-image model. The diagnostic results of the best model were compared with those of 2 radiologists.
Results:
The combined sequence model using support vector machine proves to be the best, incorporating 7 features and providing the highest values of specificity (0.903), accuracy (0.900), precision (0.727), F1score (0.800), and area under precision-recall curve (0.919) with relatively high sensitivity (0.889) in the testing set, along with a minimum Brier score. The diagnostic results differed significantly ( P < 0.05) from those of radiology residents, but not significantly ( P > 0.05) from those of experienced radiologists.
Conclusions:
Magnetic resonance imaging based on machine learning and radiomics to identify the type of sinonasal non-Hodgkin lymphoma is effective and has the potential to help radiology residents for diagnosis and be a supplement for biopsy.
