Comparing deep-learning, radiomics, and fusion models for parathyroid tumor classification using ultrasound: a multicenter retrospective study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Ultrasound, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
- School of Mathematics, Nanjing University, Nanjing, China.
- Department of Ultrasound, Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
- Department of Radiology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
- Department of Nuclear Medicine, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
Abstract
Accurate preoperative differentiation between parathyroid adenoma (PA) and parathyroid carcinoma (PC) or atypical parathyroid tumor (APT) is critical for surgical planning, yet ultrasound accuracy remains highly operator-dependent. Leveraging the complementary advantages of radiomics and deep learning (DL), this study aimed to develop and compare radiomics, DL, and fusion models based on ultrasound imaging for the identification of APT/PC. A total of 1,122 patients (270 men and 852 women; mean age 54.2±13.7 years) with parathyroid neoplasms were retrospectively reviewed from two Chinese hospitals between January 1, 2016, and April 30, 2025. To address the limited number of APT/PC cases (n=74), multicenter data were pooled and stratified by pathological type into training (733 PA and 53 APT/PC), validation (158 PA and 10 APT/PC), and test sets (57 PA and 11 APT/PC). Radiomic features were extracted from preprocessed ultrasound images. DL features came from 1ch_ResNet101 (raw images) and 2ch_ResNet101 (concatenated region of interest images). Two fusion models were built: Merged model 1 (radiomics + 1ch_ResNet101) and Merged model 2 (radiomics + 2ch_ResNet101). Statistically significant differences were observed in age at diagnosis between the training and validation sets, as well as between the validation and test sets (both P<0.001). Additionally, serum phosphate levels differed significantly between the training and test sets (P=0.022). On the test set, the top-performing models were Merged model 2 [area under the receiver operating characteristic (ROC) curve (AUC) =0.933] and Merged model 1 (AUC =0.926), both of which surpassed the ResNet101 (P<0.001) and showed comparable performance to radiomics (AUC =0.905; P=0.171 and 0.059, respectively). 2ch_ResNet101 significantly outperformed 1ch_ResNet101 (AUC =0.873 <i>vs.</i> 0.804; P<0.001). Radiomics significantly outperformed both ResNet101 models (P<0.001). No significant difference was found between the two merged models (P=0.593). All models enable timely identification of potentially malignant parathyroid tumors, with the fusion model outperforming DL and matching radiomics in diagnostic performance.