KR-20260063979-A - Auto-labeling method of object detection model

KR20260063979AKR 20260063979 AKR20260063979 AKR 20260063979AKR-20260063979-A

Abstract

The present invention relates to an auto-labeling method for an object detection model, and more specifically, to an auto-labeling method for an object detection model that is trained to detect medical diseases and utilized for auto-labeling.

Inventors

안동욱
남상도
손진호
박원형
원광재

Assignees

(주)미소정보기술

Dates

Publication Date: 20260507
Application Date: 20241031

Claims (7)

A method for auto-labeling an object detection model comprising: a first object detection step (S100) for detecting an object to be identified across the entire image of source data; a region detection step (S200) for detecting a region of an object to be identified across the entire image of source data; a second object detection step (S300) for detecting an object to be identified across an image processed from the detection result of the region detection step (S200); a learning step (S400) for learning a model using the result data of the first object detection step (S100) and the second object detection step (S300); and an auto-labeling step (S500) for utilizing the model learned through the learning step (S400) for auto-labeling.
In paragraph 1, The above first object detection step (S100) is characterized in that the object to be detected is a single object within the image of the source data or that an identification criterion exists, in an auto-labeling method of an object detection model.
In paragraph 1, An auto-labeling method for an object detection model, characterized in that a semantic consistency review by a user is performed in the first object detection step (S100) above.
In paragraph 1, An auto-labeling method for an object detection model, characterized by further including a mapping step (S600) for mapping the coordinates of an object detected in the area detection step (S200) and the second object detection step (S300) to source data.
In paragraph 1, The above second object detection step (S300) is characterized by using a processed image in which an identifiable target object appears within the entire image as input data, an auto-labeling method of an object detection model.
In paragraph 1, The above-mentioned area detection step (S200) is characterized by obtaining a correct answer sheet through user feedback on the detection result of the first object detection step (S100), reflecting it in learning, and processing the image into an area expanded to a certain level from the detected area result value and transmitting it to the second object detection step (S300), an auto-labeling method for an object detection model.
In paragraph 1, The above auto-labeling step (S500) is characterized by the detection results of the first object detection step (S100) and the second object detection step (S300) being provided simultaneously during the auto-labeling process, and the final answer sheet to be applied being selected through a review of semantic accuracy by the user.

Description

Auto-labeling method of object detection model The present invention relates to an auto-labeling method for an object detection model, and more specifically, to an auto-labeling method for an object detection model that is trained to detect medical diseases and utilized for auto-labeling. Typically, 2D or 3D medical images contain a set of anatomical and pathological structures (organs, bones, tissues, …) or artificial elements (stents, implants, …) that clinicians must depict in order to evaluate the situation and prescribe and plan their treatment. In this regard, organs and pathological findings must be identified within the images, which means labeling each pixel of the 2D image or each voxel of the 3D image. Many hospitals use a Picture Archiving Communications System (PACS) linked to medical imaging equipment such as X-ray, CT, and MRI images. The PACS stores captured medical images in digital form and transmits them to the clinician's PC, allowing the clinician to view the images on the PC and diagnose patient lesions. In particular, CT and MRI images consist of 300 consecutive cross-sectional images, placing a burden on clinicians to review all 300 images to check for lesions without omission and make an accurate diagnosis. The aforementioned CT images require a reconstruction step onto a sagittal section to visualize measurements from multiple slices obtained from a cross-sectional scan, which can be considered raw data. At this time, for the accurate analysis of lesions, various forms of CT medical images are required depending on the needs of the medical process or the equipment of each hospital, such as by having a specific sagittal slice interval or applying a specific reconstruction filter. Due to the anatomical structure of organs located deep within the human body, the function and structure of the organs themselves are complex, and measurements and examinations become complicated and imprecise due to the presence of the lungs, ribs, etc. Meanwhile, one of the key technologies in such medical imaging data is object recognition, which involves identifying and detecting target objects using an object detection model. However, there have been limitations in obtaining accurate results from medical imaging data because a single object detection model is used to identify and detect objects, despite variations in the range, resolution, or size of the object detection area of the medical imaging data. Furthermore, image data labeling is required for object recognition in medical imaging data; however, conventional labeling is mostly performed manually by operators, resulting in significant costs. Consequently, while auto-labeling methods are utilized to automatically label data using computers, they are not perfect and require manual verification. Although these conventional auto-labeling methods can improve accuracy through learning, the extent to which accuracy can be enhanced through simple learning is limited, and there is a problem in that a significant amount of time is required to improve the accuracy of the auto-labeling method. Therefore, there is a need to develop an object detection model to improve object detection accuracy, as well as a training method for said object detection model and an auto-labeling method utilizing it. FIG. 1 is a schematic flowchart of an auto-labeling method for an object detection model according to the present invention. FIG. 2 is a flowchart of an auto-labeling method of an object detection model according to the present invention including a mapping step. FIG. 3 is a flowchart illustrating the functional implementation and learning process of each object detection model in the auto-labeling method of the object detection model of the present invention. Various embodiments and/or aspects are now disclosed with reference to the drawings. For illustrative purposes, numerous specific details are disclosed in the following description to aid in a general understanding of one or more aspects. However, it will be apparent to those skilled in the art that these aspects may be practiced without such specific details. The following description and the accompanying drawings describe specific exemplary aspects of one or more aspects in detail. However, these aspects are exemplary, and some of the various methods in the principles of the various aspects may be used, and the descriptions are intended to include all such aspects and their equivalents. Specifically, terms such as “exemplary,” “example,” “aspect,” and “example” as used herein may not be interpreted as implying that any described aspect or design is superior or advantageous over other aspects or designs. Hereinafter, identical or similar components are assigned the same reference numeral regardless of drawing symbols, and redundant descriptions thereof are omitted. Furthermore, in describing the embodiments disclosed in this specification, detailed descriptions of related prior art are omitted if it is determ