Research on precise segmentation model of thyroid nodules based on deep learning

Thyroid nodules are a prevalent condition in clinical practice, typically identified via ultrasound imaging with visual assessment by physicians. However, the ambiguous features of malignant nodules, their proximity to surrounding tissues, image noise, and unclear boundaries can lead to variability in diagnostic outcomes among different practitioners. In this study, we initially employed nine established models to segment ultrasound images from two publicly available datasets, TN3K and DDTI. Evaluation metrics including IoU, F1-score, accuracy (Acc), specificity (Spec), and precision (Pre) were utilized to assess the segmentation performance of these models. The experimental results indicate that, given the current data scale, the overall segmentation performance of each model is relatively low. Analysis suggests this is primarily due to the limited number of samples in the medical imaging dataset. During training, the model struggles to fully learn the discriminative features of the lesion area. Particularly during the encoder's down-sampling stage, the feature expression ability is constrained, affecting the model's ability to accurately recognize lesion locations. To address this, we enhanced the classic model by incorporating the Residual Network ResNet34, constructing a new thyroid nodule segmentation model. The improved model demonstrated enhanced segmentation performance, with most IoU values on the TN3K dataset exceeding 0.8. However, models trained on the DDTI dataset rarely achieved high performance. Further analysis attributes this performance disparity to differences in the number of training samples. To further enhance model performance, we initialized the improved model with pre-trained ResNet34 weights from the ImageNet dataset, resulting in significant performance improvements. Notably, Res34-PAN performed best across both datasets. On the DDTI dataset, the model achieved an IoU of 0.8778, an F1-score of 0.9349, an accuracy of 0.9872, a specificity of 0.9928, and a precision of 0.9342. To further validate the model design's rationale, we conducted a systematic ablation experiment. We selected commonly used network structures for encoders and used them as the encoders for nine segmentation models, comparing performance with and without the ImageNet pre-trained weights. The results further confirmed the effectiveness of the proposed method in the thyroid nodule segmentation task.