Shareable semi-automatic diabetic fundus lesion labeling method

Abstract

The invention discloses a shareable semi-automatic diabetic fundus lesion labeling method, which comprises the following steps of identifying and segmenting the content of a diabetic fundus image based on a deep learning pre-training model, then carrying out transfer learning, and adopting different labeling methods according to different anatomical structures of fundus; carrying out semantic segmentation on blood vessels, and carrying out positioning and semantic segmentation on the optic nerve disc to output a result. For the diabetic fundus focus, positioning is adopted for hemangiomas andsemantic segmentation is adopted for exudation and bleeding so as to achieve result outputting. The contour and positioning of a focus automatically recognized through deep learning cover an originalfundus image in a layer superposition mode, and a segment with the highest confidence score is marked and presented to a marker. Manual auditing is carried out on the labeling result, and the manuallyaudited labeling data is stored by adopting a DICOM standard so as to facilitate interoperation in a standardized medical image system.

Description

technical field [0001] The invention relates to the field of artificial intelligence technology-assisted screening of fundus changes in diabetic patients, and in particular to a semi-automatic diabetic fundus lesion labeling method that can be shared using deep learning technology. Background technique [0002] The fundus can reflect the health status of the body, and it is also an important basis for judging the effect of different clinical departments on the diagnosis and treatment of patients. Therefore, in the field of medical image processing, the processing and analysis of fundus images has always been an important research content. Clinically, the judgment of abnormal fundus images requires specialized training, and there are relatively few experienced film-reading doctors, resulting in a relative shortage of medical resources. Many patients have irreversible fundus changes (diabetic retinopathy, DR) due to untimely diagnosis and treatment of the disease, and may even...

AI Technical Summary

Problems solved by technology

However, the existing automatic labeling software is mostly based on pixels, so it is not smart enough, especially on adjacent objects with similar colors.

The convolutional neural network trained by a large labeled data set can accurately identify the content on the image. However, in the field of biomedical imaging, professional skilled personnel are required to label biomedical images, which is tedious, time-consuming, and costly.

Although the annotation method based on crowdsourcing can establish a platform for a group of people to annotate large medical datasets, different annotation groups could not effectively share the results of annotations, making it very difficult to create large annotated datasets

Images