Have a personal or library account? Click to login
Dual-channel ultrasonic images empowered deep learning: significantly improving prediction of occult central lymph node metastases in solitary papillary thyroid microcarcinoma Cover

Dual-channel ultrasonic images empowered deep learning: significantly improving prediction of occult central lymph node metastases in solitary papillary thyroid microcarcinoma

Radiology and Oncology
AHEAD OF PRINT
By: Meihua Li,  Chao Jia,  Gang Li,  Qiusheng Shi and  Long Liu  
Open Access
|Feb 2026

Figures & Tables

FIGURE 1.

Flowchart depicting the construction of the dual-channel deep learning model. The model comprises two components, Model 1 and Model 2, which are built using ultrasonic images from the longitudinal and transverse sections of the thyroid tumor, respectively. These images are then fused to create dual-channel images. The ResNet101 architecture is employed to train the fused model, referred to as DL_F (Model 3). DL_C = cross-sectional deep learning model; DL_F = deep learning model using dual-channel fused images; DL_L = longitudinal deep learning model
Flowchart depicting the construction of the dual-channel deep learning model. The model comprises two components, Model 1 and Model 2, which are built using ultrasonic images from the longitudinal and transverse sections of the thyroid tumor, respectively. These images are then fused to create dual-channel images. The ResNet101 architecture is employed to train the fused model, referred to as DL_F (Model 3). DL_C = cross-sectional deep learning model; DL_F = deep learning model using dual-channel fused images; DL_L = longitudinal deep learning model

FIGURE 2.

Flowchart illustrating the patient enrollment process. PTC = papillary thyroid carcinoma; PTMC = papillary thyroid microcarcinoma; CLNM = cervical lymph node metastasis
Flowchart illustrating the patient enrollment process. PTC = papillary thyroid carcinoma; PTMC = papillary thyroid microcarcinoma; CLNM = cervical lymph node metastasis

FIGURE 3.

Gradient-weighted Class Activation Mapping images predicted by the dual-channel deep learning model. The left image represents the minimum bounding rectangle containing the largest section of the tumor on the longitudinal section. The middle image represents the minimum bounding rectangle containing the largest section of the tumor on the transverse section. The right image represents the Gradient-weighted Class Activation Mapping image.
Gradient-weighted Class Activation Mapping images predicted by the dual-channel deep learning model. The left image represents the minimum bounding rectangle containing the largest section of the tumor on the longitudinal section. The middle image represents the minimum bounding rectangle containing the largest section of the tumor on the transverse section. The right image represents the Gradient-weighted Class Activation Mapping image.

FIGURE 4.

The AUC values of the combined model. Compared with the clinical model and the dual-channel model, the combined model has higher AUC values on the training set (A), the validation set (B), and the external test set (C). AUC = area under the receiver operating characteristic curve; CI = confidence interval; LightGBM = Light Gradient Boosting Machin; LR = logistic regression
The AUC values of the combined model. Compared with the clinical model and the dual-channel model, the combined model has higher AUC values on the training set (A), the validation set (B), and the external test set (C). AUC = area under the receiver operating characteristic curve; CI = confidence interval; LightGBM = Light Gradient Boosting Machin; LR = logistic regression

FIGURE 5.

Calibration curves of different models. (A) Calibration curves of the combined model, dual-channel model and the clinical model in training dataset; (B) Calibration curves in the validation dataset; (C) Calibration curves in the external test set.
Calibration curves of different models. (A) Calibration curves of the combined model, dual-channel model and the clinical model in training dataset; (B) Calibration curves in the validation dataset; (C) Calibration curves in the external test set.

FIGURE 6.

Contributions of individual features in the combined model. The summary plot displays the distribution of SHAP values for each feature across all predictions in the training cohort, providing a comprehensive overview of feature importance and their respective impacts. SHAP = SHapley Additive exPlanations
Contributions of individual features in the combined model. The summary plot displays the distribution of SHAP values for each feature across all predictions in the training cohort, providing a comprehensive overview of feature importance and their respective impacts. SHAP = SHapley Additive exPlanations

Characteristics of the distribution of clinical data of patients among groups

Clinical indicatorsTraining set (n = 265)Validation set (n = 114)External test set (n = 82)StatisticsP value
Age (years) a43.0 (35.0–52.0)45.5 (35.3–54.0)45.0 (34.0–53.0)0.4010.818
Size (mm) a6.1 (4.7–7.8)5.6 (4.2–7.0)6.0 (4.6–7.3)4.0900.129
Aspect ratio a1.1 (0.9–1.3)1.0 (0.9–1.2)1.1 (0.9–1.3)3.5470.170
Sex 1.4930.474
     Male69 (26.0)24 (21.1) b23 (28.1) b
     Female196 (74.0)90 (79.0)59 (72.0)
Location 15.6320.016
     Upper segment33 (12.5)15 (13.2)22 (26.8) b
     Middle segment146 (55.1)58 (50.9)38 (46.3)
     Lower segment57 (21.5)29 (25.4)20 (24.4)
     Isthmus29 (10.9)12 (10.5)2 (2.4)
ETE 4.7330.094
     Absence226 (85.3)102 (89.5)77 (93.9)
     Presence39 (14.7)12 (10.5)5 (6.1)
Echogenicity 4.1100.391
     Hyperechoic/isoechoic11 (4.2)8 (7.0)5 (6.1)b
     Hypoechoic200 (75.5)75 (65.8)59 (72.0)
Markedly
     Hypoechoic54 (20.3)31 (27.2)18 (22.0)
Margin 2.4490.294
     Clear113 (42.6)53 (46.5)29 (35.4)
     Unclear152 (57.4)61 (53.5)53 (64.6)
Shape 5.2420.073
     Regular115 (43.4)47 (41.2)24 (29.3)
     Irregular150 (56.6)67 (58.8)58 (70.7)
Vascularity 5.6420.465
     Type I95 (35.9)38 (33.3)24 (29.3)
     Type II21 (7.9)9 (7.9)12 (14.6)
     Type III140 (52.8)63 (55.3)41 (50.0)
     Type IV9 (3.4)4 (3.5)5 (6.1)
Calcification Absence123 (46.4) b58 (50.9)41 (50.0)2.9020.821
     Microcalcification100 (37.7)38 (33.3)25 (30.5)
     Coarse type25 (9.4)13 (11.4)10 (12.2)
     Mixed type17 (6.4)5 (4.4)6 (7.3)
CC 0.2120.900
     Absence97 (36.6)41 (36.0)32 (39.0)
     Presence168 (63.4)73 (64.0)50 (61.0)
DC 0.4960.780
     Absence186 (70.2)83 (72.8)56 (68.3)
     Presence79 (29.8)31 (27.2)26 (31.7)
DOI: https://doi.org/10.2478/raon-2026-0006 | Journal eISSN: 1581-3207 | Journal ISSN: 1318-2099
Journal RSS Feed
Language: English
Submitted on: May 5, 2025
|
Accepted on: Aug 20, 2025
|
Published on: Feb 6, 2026
Published by: Association of Radiology and Oncology
In partnership with: Paradigm Publishing Services
Publication frequency: 4 issues per year
Keywords:
papillary thyroid microcarcinoma,
central lymph node metastasis,
deep learning,
ultrasound
Related subjects:
Medicine,
Clinical medicine,
Internal medicine,
Haematology, oncology,
Radiology

© 2026 Meihua Li, Chao Jia, Gang Li, Qiusheng Shi, Long Liu, published by Association of Radiology and Oncology
This work is licensed under the Creative Commons Attribution 4.0 License.

AHEAD OF PRINT