Related Experiment Video For Deep learning
Updated: Jan 7, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Automated segmentation and diagnosis of parotid tumors using a combined deep learning and radiomics model on
Qifan Ma1, Jiliang Ren2, Yaqiong Ge3
1Department of Radiology, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, No. 639 Zhizaoju Road, Shanghai, 200010, China.
Objectives:
To develop and validate an automated diagnostic framework that combines deep learning and radiomics models for the segmentation and classification of benign and malignant parotid gland tumors on magnetic resonance imaging (MRI).
Methods:
In total, 493 patients with pathologically confirmed parotid tumors (396 benign and 97 malignant) were included. Patients were stratified by MRI scanner type into a training cohort (n = 288), an internal validation cohort (n = 123), and an external testing cohort (n = 82). An automated tumor segmentation model based on the nnU-NetV2 architecture was developed and evaluated using the Dice similarity coefficient (DSC) and Intersection over Union (IoU). Based on the automatically segmented regions, a radiomics-based classifier and a ResNet18-based deep learning model were independently constructed to differentiate malignant from benign tumors. A combined diagnostic model was further developed by integrating deep learning outputs, radiomics features, and clinical-radiological features. Model performance was assessed using the area under the receiver operating characteristic curve (AUC).
Results:
The automated segmentation model achieved a Dice similarity coefficient (DSC) of 0.93 and an Intersection over Union (IoU) of 0.88 in the training cohort, 0.91 and 0.84 in the validation cohort, and 0.84 and 0.76 in the testing cohort, respectively. The ResNet18-based DL model achieved AUCs of 0.90, 0.84, and 0.77, respectively, compared to the radiomics model's AUCs of 0.79, 0.72, and 0.71. The combined model demonstrated superior performance, with AUCs of 0.92 in the validation cohort and 0.90 in the testing cohort, outperforming the clinical-radiological model, which achieved AUCs of 0.69 and 0.82 (p < 0.001 in validation, p = 0.042 in testing).
Conclusions:
This automated MRI-based framework, combining deep learning and radiomics approaches, enables accurate segmentation and reliable classification of parotid gland tumors. It offers a promising noninvasive tool to assist in clinical decision-making.
Clinical Trial Number:
Not applicable.
