Improved Deeplabv3+ based ultrasound image segmentation algorithm for thyroid nodules

Abstract

The incidence of thyroid nodules is increasing year by year, and ultrasonic detection is an effective means of thyroid nodule lesions, but the background of ultrasound images is chaotic, and the lesion area is similar to the background. The traditional segmentation algorithm produces a large number of invalid feature channels in the stage of encoding feature extraction, and the network continuously downsamples and pooling operations make the detail information of the thyroid nodule edge under the ultrasound image lose the detail information and the malignant nodule edge segmentation effect is not good. In view of the above problems, this paper proposes a thyroid nodule ultrasound image segmentation algorithm based on improved Deeplabv3+, which integrates the attention mechanism on the original Deeplabv3+ network structure, and designs a new attention mechanism module ES-Net, which adds a spatial attention module on the basis of ECA-Net channel attention, which can effectively pay attention to the spatial structure under the ultrasound image. Experimental data show that in the public dataset of thyroid ultrasound images, the Dice loss decreased from 0.192 to 0.120, and the IoU (Intersection of Union) increased from 87.80% to 89.62%. The results show that the improved Deeplabv3+ attention model has a 1.82% accuracy improvement in the segmentation of ultrasound images of thyroid nodules, which verifies the effectiveness of the algorithm.

Ähnliche Arbeiten