Doppler-Enhanced Deep Learning: Improving Thyroid Nodule Segmentation with YOLOv5 Instance Segmentation

• Ultrasound detects thyroid nodules frequently, but only 5–15% of sonographically detected nodules are malignant, so distinguishing those requiring surgery from benign nodules is the central clinical challenge.Source: §1 · p. 1
• TIRADS systems still have 49–56% false-positive biopsies and rely on qualitative interpretation subject to expert inter-observer variability (Kappa=0.519).Source: §1 · p. 1
• Accurate thyroid nodule segmentation is a crucial first step toward AI-powered clinical decision support for thyroid cancer detection and diagnosis.Source: §1 · p. 2
• The study evaluates YOLOv5 models for automatic thyroid nodule instance segmentation, reporting high accuracy and computational efficiency on a thyroid ultrasound dataset.Source: §1 · p. 2

2.1 Deep Learning for Medical Image Segmentation

• Traditional feature engineering struggled with ultrasound noise and variability, while CNN-based U-Net models became common but may not distinguish individual instances of the same class.Source: §2.1 · p. 2

2.2 Instance Segmentation Approaches

• Mask R-CNN combines object detection with pixel-level masks; Abdolali et al. reported their thyroid-nodule framework outperformed Faster R-CNN and Mask R-CNN.Source: §2.2 · p. 2

2.3 YOLO-based Approaches

• YOLO offers single-stage real-time detection, and prior improved YOLOv5 work reported competitive thyroid-nodule detection with reduced computational requirements.Source: §2.3 · p. 2

2.4 Research Gap

• Systematic comparisons of YOLOv5 variants for thyroid nodule instance segmentation, especially with Doppler imaging, remain limited; this study evaluates multiple models and dataset compositions.Source: §2.4 · p. 2

3.1 Dataset

• V1 contains 2,075 patients, 4,129 images, and 4,407 nodules with Doppler images; V2 contains 1,969 patients, 3,932 images, and 4,192 nodules without them.V2 excludes 197 Doppler-elastography images; images labeled invalid or containing artifacts were also excluded.Source: §3.1 · p. 3

3.2 Data Preprocessing

• Raw DICOM images were converted to normalized grayscale PNGs, and polygon nodule regions were extracted into masks and converted to YOLO and COCO formats.Source: §3.2 · p. 3

3.3 Instance Segmentation with YOLOv5

• YOLOv5-seg adds a segmentation branch to predict masks alongside boxes and labels in one forward pass; five model sizes were tested, with Large offering the best speed–accuracy trade-off.The variants range from Nano (2.0M parameters) to XLarge (88.8M parameters).Source: §3.3 · p. 4

3.4 Training Configuration

• Data were split 80%/15%/5% for training, validation, and testing without augmentation; dropout and weight decay were omitted because validation performance showed no overfitting.Hyperparameters were iteratively adjusted to balance training speed and model performance.Source: §3.4 · p. 4

3.5 Evaluation Metrics

• Evaluation used Dice overlap, precision for correct detections, recall for detected instances, and mAP@0.5 based on an IoU threshold of 0.5.Source: §3.5 · pp. 4–5

4 Results

• Table 1 covers hyperparameter configurations for YOLOv5 variants spanning 2.0M parameters for Nano to 88.8M for XLarge.The configurations provide comparisons across different model complexities.Source: §4 · p. 5, Table 1 · p. 5

4.1 Performance Analysis

• Including Doppler improved scores across model variants: Large rose from 0.76 on V2 to 0.91 on V1, a 19.7% improvement.The authors report consistent improvement across model sizes and say Doppler provides complementary information for nodule boundary delineation.Source: §4.1 · pp. 5–6
• YOLOv5-Large achieved a 91% test dice score and 0.87 mAP on V1; Medium scored 0.90 dice, while XLarge scored 0.89.Large used 47.9M parameters, compared with 22.0M for Medium and 88.8M for XLarge.Source: Table 2 · p. 5, §4.1 · p. 6
• Nano achieved a 0.84 dice score on V1 with 2.0M parameters; Small scored 0.87 with 7.6M parameters, using 84% fewer parameters than Large.The Small model's performance reduction relative to Large was 4.4%.Source: §4.1 · p. 6
• Precision exceeded recall across all models, indicating more conservative predictions that favored avoiding false positives over detecting every nodule.The paper describes this as clinically favorable because false positives can prompt anxiety and additional examinations, while missed nodules may be detected in follow-up screenings.Source: §4.1 · p. 6
• YOLOv5-Large segmented four distinct nodules in one ultrasound image, with confidence scores above 0.85 across detections.The authors present multi-nodule detection as relevant to images containing nodules of varying sizes.Source: Figure 1 · p. 6, §4.1 · p. 6

5 Discussion

• YOLOv5-Large achieved a 91% dice score, supporting the potential of real-time automated thyroid nodule detection.The authors also describe automated segmentation as a possible second reader or training aid for less experienced clinicians.Source: §5 · p. 7
• Doppler images significantly improved automated segmentation, and vascularization patterns may help delineate nodule boundaries.The authors suggest retaining Doppler images in training datasets and reconsidering them for clinical evaluation.Source: §5 · p. 7
• YOLOv5-Small (7.6M parameters) achieved 87% dice score on V1, while inference typically took under 50ms per image on modern GPUs.The authors present smaller variants as options for mobile devices or systems with limited computational resources.Source: §5 · p. 7
• Despite dataset differences limiting direct comparison, YOLOv5-Large’s 91% dice score was competitive with or superior to reported Mask R-CNN results.The authors attribute a real-time advantage to YOLOv5’s single-stage architecture compared with two-stage detectors.Source: §5 · p. 7
• With approximately 4,100 training images, YOLOv5-XLarge (88.8M parameters) performed slightly worse than YOLOv5-Large (47.9M parameters).The authors suggest model capacity should be matched to dataset size and that larger datasets may enable effective use of larger models.Source: §5 · p. 7

5.1 Limitations

• The single-institution dataset and static-image evaluation constrain generalizability across clinical settings and do not assess temporal variation in sequential imaging or cine-loops.Manual-label inter-observer variability was not quantified, and data augmentation was not applied.Source: §5.1 · p. 7
• Performance was not stratified by nodule size, TIRADS classification, or malignancy status, and inference costs were not systematically measured across hardware platforms.Source: §5.1 · p. 8

6 Conclusion

• YOLOv5-Large achieved a 91% dice score and mAP of 0.87 with Doppler images, which improved performance across variants by 8.9% to 19.7%.Source: §6 · p. 8
• The evaluated YOLOv5 variants range from lightweight models suited to mobile devices to larger models optimized for accuracy.The authors conclude that real-time, accurate automated segmentation is achievable with current deep learning technology.Source: §6 · p. 8
• YOLOv5-Medium (22.0M parameters) achieved a 90% dice score, which the authors describe as near-optimal performance for resource-constrained clinical environments.Source: §6 · p. 8