Background
Objective
Methods
Image Collection
-
Database 1 is a publicly available thyroid ultrasound image database proposed by Pedraza et al. [16], consisting of 428 thyroid ultrasound images with the size 560 × 360, of which 357 cases are labelled as positive (with the TI-RADS scores 3, 4a, 4b or 5), while 71 cases are labelled as negative (with the TI-RADS scores 1 or 2). The images were extracted from thyroid ultrasound video sequences captured with a TOSHIBA Nemio 30 and a TOSHIBA Nemio MX Ultrasound devices, both set to 12 MHz convex and linear transductors, containing the most relevant pathological features and their pathologies confirmed by biopsy using the BETHESDA system. The experts independently evaluated the patient individually and described the specific features filling the TI-RADS requirements;
-
Database 2 is a local database, consisting of 164 thyroid ultrasound images, of which 122 images are with the sizes 1024(±5) × 695(±5), while 42 images are with the sizes 640(±5) × 440(±5). Thirty-five cases of the images in the database are labelled as positive (with TI-RADS scores of 3, 4a, 4b or 5), while 129 cases are labelled as negative (with TI-RADS scores of 1 or 2). The TI-RADS scores were evaluated by the local experts using the same method as those in database 1.
Image Pre-processing
Sample Augmentation
Total | Positive | Negative | ||||
---|---|---|---|---|---|---|
Cases | Samples | Cases | Samples | Cases | Samples | |
Training | 306 | 2754 | 256 | 2304 | 50 | 450 |
Validating | 61 | 549 | 51 | 459 | 10 | 90 |
Testing | 61 | 549 | 50 | 450 | 11 | 99 |
Total | Positive | Negative | ||||
---|---|---|---|---|---|---|
Cases | Samples | Cases | Samples | Cases | Samples | |
Training | 132 | 1188 | 21 | 189 | 111 | 999 |
Validating | 16 | 144 | 4 | 36 | 12 | 108 |
Testing | 16 | 144 | 4 | 36 | 12 | 108 |
Fine-Tuned GoogLeNet Model Generation
Feature Extraction and Classification Model
Model Assessment
Results
Model Selection
Model Assessment
Accuracy | Sensitivity | Specificity | AUC | |
---|---|---|---|---|
Proposed model | 99.13% | 99.70% | 95.80% | 0.9970 |
Accuracy | Sensitivity | Specificity | AUC | |
---|---|---|---|---|
Proposed model | 96.34% | 82.80% | 99.30% | 0.9920 |
Discussion
Method | Feature | Machine learning | Testing samples | Accuracy |
---|---|---|---|---|
Lim et al. [24] | Size, margin, echogenicity, cystic change | Artificial neural network | 190 thyroid lesions | 93.78% |
Savelonas et al. [25] | Shape features | Nearest neighbour (k-NN) | 173 longitudinal in vivo images | 93.00% |
Iakovidis et al. [26] | Fuzzy intensity histogram | SVM | 250 thyroid ultrasound images | 97.50% |
Legakis et al. [27] | Texture features, shape features | SVM | 142 longitudinal in vivo images | 93.20% |
Luo et al. [28] | Strain rate waveform’s power spectrum | Linear discriminant analysis | 98 nodule image sequences | 87.50% |
Acharya et al. [7] | Higher-order spectra features | Fuzzy classifier | 80 sample 3D images | 99.10% |
Proposed model | Deep learning features | Cost-sensitive Random Forest classifier | 693 sample images | 99.13% |