Tnseg: adversarial networks with multi-scale joint loss for thyroid nodule segmentation

The thyroid gland is a critical regulator of numerous physiological functions, and the presence of thyroid nodules often signals potential disorders. Accurate nodule segmentation from ultrasound images is imperative for effective diagnosis and treatment planning. Existing techniques often struggle due to intra-nodule variability. To address this, we introduce TNSeg, an innovative framework specifically designed for thyroid nodule segmentation. TNSeg incorporates two key components: a segmentation block and a discriminative block, and leverages adversarial training. In particular, the discriminator uses a fully convolutional decoder with skip connections to efficiently differentiate between real and synthetic samples. Further, we introduce a novel multi-scale joint loss function for adversarial training that employs a balanced sampling strategy, effectively resolving the difficulties associated with foreground-background differentiation and computational redundancy. Extensive evaluation proves TNSeg’s superiority in achieving a Dice coefficient of 92.06%, Hd95 of 13.35, Jaccard index of 90.02%, and Precision of 94.01%, thereby demonstrating significant improvements in four commonly used segmentation quality metrics.