Method and apparatus for selecting medical data for annotation
An operating method of a medical data selecting apparatus operated by at least one processor includes generating training data including partial medical data sampled from mass medical data and annotated data of the partial medical data, extracting candidate data for annotation from the mass medical data, the candidate data being at least a portion of the mass medical data, acquiring inference results that are inferred from the candidate data by an artificial intelligence (AI) model trained based on the training data and selecting target data for annotation to be used in next training of the AI model, from among the candidate data based on the inference results.
1 . An operating method of a medical data selecting apparatus operated by at least one processor, the operating method comprising:
generating training data including partial images sampled from mass images corresponding to medical data and annotated data of the partial images;
extracting candidate images for annotation from the mass images, the candidate images being at least a portion of the mass images and being images to which annotated data is not mapped;
acquiring inference results that are inferred from the candidate images by an artificial intelligence (AI) model trained based on the training data; and
selecting target images for annotation to be used in next training of the AI model, from among the candidate images based on the inference results;
performing the next training of the AI model based on the target images and annotated data of the target images,
wherein the selecting the target images for annotation includes:
determining a selection policy for images to be used in the next training based on training performance of the AI model that has been trained in previous training; and
selecting the target images for annotation corresponding to the selection policy from among the candidate images based on the inference results.
2 . The operating method of claim 1 , wherein the selection policy includes a policy of extracting images classified in a specific class more than images classified in other classes, the specific class having a classification accuracy equal to or lower than a reference.
3 . The operating method of claim 1 , further comprising:
extracting validation data from the mass images, the validation data being at least a portion of the mass images; and
evaluating the training performance of the AI model based on the validation data.
4 . The operating method of claim 1 , wherein the selecting the target images for annotation includes adding images randomly extracted from the mass images to the target images for annotation.
5 . The operating method of claim 1 , wherein the selecting the target images for annotation includes
selecting images related to a specific class as the target images from among the candidate images based on the inference results.
6 . The operating method of claim 1 , wherein the generating the training data includes:
analyzing a reading report associated with the mass images;
sampling the partial images associated with the reading report including information related to training of the AI model; and
acquiring the annotated data of the partial images to generate the training data.
7 . The operating method of claim 1 , wherein the mass images include images acquired by at least one medical imaging device, pathological images, or patch images extracted from a medical image.
8 . An operating method of a medical data selecting apparatus operated by at least one processor, the operating method comprising:
repeating a process including:
determining a selection policy for images corresponding to medical data based on training performance of a current artificial intelligence (AI) model,
selecting annotation target images to be used in next training corresponding to the selection policy from among candidate images for annotation, based on inference results that are inferred from the candidate images by the current AI model, the candidate images being at least a portion of the mass images and being images to which annotated data is not mapped, and
training the current AI model based on annotated data acquired for the annotation target images; and
terminating the process when the next training of the current Al model is not performed,
wherein the selecting the annotation target images includes:
determining a selection policy for images based on training performance of the current AI model that has been trained in previous training; and
selecting the annotation target images corresponding to the selection policy from among the candidate images based on the inference results.
9 . The operating method of claim 8 , wherein the repeating includes:
extracting validation data from the mass images, the validation data being at least a portion of the mass images; and
evaluating the training performance of the current AI model based on the validation data.
10 . The operating method of claim 8 , wherein the process further includes adding images randomly extracted from the mass images to the annotation target images.
11 . The operating method of claim 8 , wherein the process further includes determining the selection policy to be a same as or different from a selection policy for the previous training based on the training performance of the current AI model.
12 . The operating method of claim 8 , wherein the current AI model includes an initial model trained based on partial images sampled from the mass images and annotated data of the partial images, or a model obtained by retraining the initial model based on the process.
13 . The operating method of claim 8 , wherein the mass images include images acquired by at least one medical imaging device, pathological images, or patch images extracted from a medical image.
14 . A medical data selecting apparatus, the medical data selecting apparatus comprising:
an artificial intelligence (AI) model trained to output an inference result that is inferred from an input based on training data; and
at least one processor configured to:
extract candidate images for annotation from mass images corresponding to medical data, the candidate images being at least a portion of mass images and being images to which annotated data is not mapped,
acquire inference results that are inferred from the candidate images by the AI model,
select annotation target images to be used in next training of the AI model from among the candidate images based on the inference results,
acquire annotated data of the annotation target images to be used in the next training of the AI model, and
perform the next training of the AI model based on the annotation target images and the annotated data,
wherein the at least one processor is further configured to:
determine a selection policy for images to be used the next training based on training performance of the AI model that has been trained in previous training; and
select the annotation target images corresponding to the selection policy from among the candidate images based on the inference results.
15 . The medical data selecting apparatus of claim 14 , wherein the at least one processor is further configured to repeat a process of selecting the annotation target images to be used in the next training based on the inference results of the AI model when the next training of the AI model is required.
16 . The medical data selecting apparatus of claim 14 , wherein the at least one processor is further configured to determine the selection policy to be a same as or different from a selection policy for the previous training based on the training performance of the AI model.
17 . The medical data selecting apparatus of claim 14 , wherein the at least one processor is further configured to add images randomly extracted from the mass images to the annotation target images.
18 . The medical data selecting apparatus of claim 14 , wherein the mass images include images acquired by at least one medical imaging device, pathological images, or patch images extracted from a medical image.