An Efficient Multi-Task FCOS Network Based on MobileNetV3 for Thyroid Nodule Localization and Classification
Abstract
Accurately diagnosing thyroid nodules via ultrasound poses challenges due to blurred boundaries and a heavy reliance on physician experience, while existing high-precision deep learning models are often difficult to deploy on edge devices due to their computationally intensive nature. To address this, this paper proposes a lightweight, anchor-free Multi-Task Learning (MTL) framework based on MobileNetV3-Small. By combining a Feature Pyramid Network (FPN) with a single-target decoding strategy, the model achieves joint optimization of nodule benign/malignant classification and precise localization. Additionally, the introduction of DropBlock2D regularization in the FCOS detection head effectively mitigates local overfitting caused by ultrasound speckle noise. Five-fold cross-validation on the TN5000 dataset demonstrates that the model achieves excellent classification (AUC: 0.9406) and detection (mAP@0.50: 0.8716) performance while maintaining an extremely low computational footprint (only 1.065 M parameters and 0.640 G FLOPs), and accurately localizes lesions in 91.3% of the samples. Furthermore, qualitative analysis confirms the framework’s advantage in identifying potential multiple missed lesions. The method presented in this paper achieves an excellent trade-off between pattern recognition accuracy and computational complexity, providing an efficient and lightweight solution for biomedical image processing and intelligent visual analysis systems.