The application relates to a
medical ultrasound image dataset construction method, which comprises the following steps: acquiring an
ultrasound image for labeling; training a multi-
modal segmentation model based on the
labeled data; acquiring a to-be-labeled
ultrasound image, calculating the similarity between the to-be-labeled
ultrasound image and the labeled
ultrasound image to select a
reference image; performing weighted average on the
mask corresponding to the
reference image based on the similarity to obtain a fusion
mask; performing weighted fusion on the circumscribed rectangular region corresponding to the
mask of the
reference image based on the similarity to obtain a bounding box prompt; taking the to-be-labeled
ultrasound image, the corresponding fusion mask, a text
label and the bounding box prompt as inputs of the multi-
modal segmentation model to generate a segmentation result, adding the segmentation result to a dataset after confidence evaluation, and performing subsequent model training and reference
image selection. Compared with the prior art, the application significantly reduces the
workload of manual labeling, reduces the professional requirements for personnel, and greatly improves the
data set labeling efficiency and the manual correction speed.