A novel network for resolving subjective masking differences and accurate thyroid nodule diagnosis.

Background: Over the past three decades, there has been a significant increase in the incidence of thyroid cancer. Ultrasound serves as a non-invasive tool in differentiating between benign and malignant thyroid nodules. However, its reliance on manual input can often lead to subjective bias.

Purpose: This study proposes a novel network architecture committed to diminishing subjective bias led by manual masks and enhancing the accuracy of the current models. It amalgamates multi-scale features for the effective classification of thyroid nodules.

Methods: The innovative model, deemed APSNet, finds inspiration from active and passive systems. It incorporates attention mechanisms to augment nodule recognition. The model underwent training on a localized ultrasound image dataset and was tested using an external datasets TDID and TN3K. The assessment of its performance involved metrics such as Dice, IoU, F1, Acc, Sen, Spe, Ppv, Npv, and AUC, followed by statistical tests including the Friedman and DeLong tests.

Results: APSNet outperformed existing models across multiple metrics, achieving an Acc of 0.9259, F1 score of 0.9540, and AUC of 0.9243 on the TDID dataset, and an Acc of 0.9287, F1 score of 0.9001, sensitivity of 0.9273, and AUC of 0.9290 on the TN3K dataset. The DeLong test confirmed its superiority, indicating statistically significant improvements over other models. Ablation Study confirms the effectiveness of Dual-System design and the potention of Transformer-based backbone.

Conclusions: APSNet offers a remarkable stride forward in thyroid nodule diagnosis by effectively addressing subjectivity and amplifying feature extraction capabilities. It proffers a more accurate and dependable diagnostic tool to clinicians.