Laser & Optoelectronics Progress, Volume. 61, Issue 8, 0800002(2024)
Application Progress of Deep Learning in the Classification of Benign and Malignant Thyroid Nodule
Figures & Tables(6)
![Detection process[35]](/Images/icon/loading.gif){kind=icon}
Table 1. Classification performance evaluation index and their mathematical description
View tableTable 1. Classification performance evaluation index and their mathematical description
Evaluation index Mathematical description ACC TPR TNR PPV Recall F1-score
Table 2. Summary of TN classification methods based on deep learning
View tableTable 2. Summary of TN classification methods based on deep learning
Method Main idea Advantage Disadvantage Single CNN Extract image features and minimize loss values Low complexity，short computation time，and relatively easy to improve Incomplete feature extraction，limited recognition ability Multiple CNN Integrate multiple CNN models and leverage multiple underlying network architectures Capture global and local feature information of nodules，with stronger feature expression ability and higher classification accuracy Multiple parameters，difficult training，and long calculation time Transformer Serializing images and utilizing self attention mechanism to capture global dependencies of images Capture global dependencies，parallel computing，long-distance dependency modeling，and strong interpretability High computational resource requirements，loss of image structure information，and more complex feature processing DNN Using deep neural networks for feature transformation and complex function fitting Enhance diagnostic homogeneity，quantitatively extract image features，and output conclusions through standardized processing methods Difficulty in training and low model specificity GAN Continuously adversarial image generation through generators and discriminators Enhance training stability，solve the problem of low manually labeled data volume，and increase the number and diversity of training samples Difficulty in model training，problem of pattern collapse Transfer learning Migrating annotation data from related auxiliary fields Enhance generalization ability and alleviate training difficulties caused by the lack of medical images in the model The migration layer and migration volume need to be verified，have negative migration issues Ensemble learning Use ensemble learning to combine decisions from multiple CNN Higher classification accuracy and enhanced prediction stability Complex network，difficult training，and long training time
Table 3. Performance comparison of TN classification methods based on deep learning
View tableTable 3. Performance comparison of TN classification methods based on deep learning
Classification method ACC /% TPR /% TNR /% CNN+Transfer learning+Mixup［ 32 ］90.15 97.40 87.30 ResNet-18［ 34 ］98.40 97.80 — TV+GoogLeNet［ 35 ］96.04 — — Inception-ResNet-v2+AlexNet［ 36 ］87.32 84.22 — DMRF-CNN［ 39 ］95.24 — 97.39 ResNet-50+Inception［ 42 ］92.05 96.07 65.69 Transfer learning+Inception V3［ 67 ］92.85 — — Transfer learning+DenseNet161［ 67 ］92.91 — — Ensemble learning+EDLC-TN+DenseNet［ 79 ］95.76 95.88 93.75 VGG-16+BN［ 33 ］86.43 87.43 85.43 EfficientNet B7+MBConv［ 37 ］92.80 — 96.30 EDSResNet［ 41 ］92.40 94.50 91.70 IF-JCNN［ 46 ］89.60 88.50 91.00 AlexNet+VGG-16+ResNet-50+Transfer learning ［ 49 ］96.00 94.10 97.70 5-CNN+VGG-19［ 53 ］95.70 98.43 97.35 Res-GAN［ 59 ］95.00 — — DNN+DNN+DenseNet［ 62 ］— 88.10 83.90 Transfer learning+TV+GoogLeNet［ 68 ］96.04 — — Ensemble learning+Transfer learning+CNN［ 73 ］86.70 84.70 — ResNet-50+ResNet-50［ 43 ］92.76 95.68 84.75 CAM+GAN［ 58 ］82.81 — — SCGAN+ResNet-50［ 60 ］94.30 — — Transfer learning+ResNet［ 44 ］88.30 — — Vision Transformer［ 54 ］86.90 — — Transfer learning+ResNet-50+ResNet-50［ 69 ］90.34 — — Ensemble learning+DenseNet121+ResNet-50+
Inception V3+LSTM+Conv4［
80 ］91.90 97.40 —
Table 4. Performance comparison of TN CAD system based on deep learning
View tableTable 4. Performance comparison of TN CAD system based on deep learning
CAD system ACC /% TPR /% TNR /% AUC ResNet-50［ 86 ］90.40 90.40 90.40 0.884 AI-Sonic CAD［ 87 ］76.10 — 93.00 0.919 HT-CAD［ 88 ］87.60 90.50 83.00 0.867 CAD［ 89 ］— 90.70 74.60 0.830 S-Detect［ 90 ］— 90.50 — — VGG-16+RNN［ 91 ］99.95 99.95 — 1.000
Tools
Get Citation
Copy Citation Text
Wenkai Zhang, Xiaoyan Wang, Jing Liu, Qixiang Zhou, Xin He. Application Progress of Deep Learning in the Classification of Benign and Malignant Thyroid Nodule[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0800002
Paper Information
Category: Reviews
Received: Jun. 5, 2023
Accepted: Aug. 1, 2023
Published Online: Mar. 1, 2024
The Author Email: Wang Xiaoyan (sdnuwxy@126.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20