Advanced Ultrasound in Diagnosis and Therapy ›› 2018, Vol. 2 ›› Issue (2): 82-93.doi: 10.37015/AUDT.2018.180804

• Original Research • Previous Articles     Next Articles

Deep Learning Models for Segmentation of Lesion Based on Ultrasound Images

Jinlian Ma, PhDa, Dexing Kong, PhDb,*()   

  1. a State Key Lab of Computer-Aided Design & Computer Graphics, College of Computer Science and Technology, Zhejiang University, China
    b School of Mathematical Sciences, Zhejiang University, China
  • Received:2018-07-02 Online:2018-08-18 Published:2018-08-19
  • Contact: Dexing Kong, PhD,


Objective: Ultrasonography is widely used for the diagnosis of many diseases including thyroid and breast cancers. Delineation of lesion boundaries from ultrasound images plays an important role in calculation of clinical indices and early diagnosis of diseases. However, accurate automatic segmentation of lesions is challenging because of their heterogeneous appearance and lack of background contrast.
Method: In this study, we employed a pre-trained deep convolutional neural network (PCNN) to automatically segment lesions from ultrasound images. Specifically, our PCNN based method used whole images of normal tissues and lesions as inputs and then generated the segmentation probability maps as outputs. A pre-training strategy was used to improve the performance of the PCNN based model. Additionally, we compared the performance of our approach with that of the common convolutional neural network segmentation methods on the same dataset.
Results: We validated on ultrasound images of thyroid nodules and breast nodules. The experimental results were shown in true positive rate (TP), false positive rate (FP), overlap metric (OM) and dice ratio (DR). Specifically, for thyroid nodule segmentation, our method could achieve an average of OM, DR, TP, FP as 0.8943, 0.9558, 0.9694, 0.0569 on overall folds, respectively. For breast nodule segmentation, our method could achieve an average of OM, DR, TP, FP as 0.8572, 0.9001, 0.9497, 0.8619, respectively.
Conclusion: Our proposed method is fully automatic without any user interaction and may be good enough to replace the timeconsuming and tedious manual segmentation approach, demonstrating the potential clinical applications.

Key words: Ultrasound image; Convolutional neural network; Pre-training; Segmentation