EP-4502830-B1 - METHOD FOR FILTERING NORMAL MEDICAL IMAGE, METHOD FOR INTERPRETING MEDICAL IMAGE, AND COMPUTING DEVICE IMPLEMENTING THE METHODS

Inventors

• Park, Jongchan

Dates

Publication Date

20260513

Application Date

20210609

Claims (9)

1. A computing device (10) comprising: a memory; and at least one processor (11) that executes instructions stored in the memory, wherein the processor (11) is configured to: obtain disease prediction scores for different diseases predicted in a medical image using abnormality prediction model (100); determine abnormality scores for the different diseases using the disease prediction scores for the different diseases; calculate cut-off scores for the different diseases which makes a specific reading sensitivity; classify the medical image into strong normal or not strong normal using the abnormality scores for the different diseases; and add the medical image to a reading worklist depending on the classification result, wherein, in order to perform the step of classifying the medical image, the processor (11) is configured to: classify the medical image into the strong normal when the abnormality scores for the different diseases are less than or equal to the cut-off scores corresponding to the different diseases, for all of the different diseases, wherein the medical image classified into strong normal is excluded from the reading working list.
2. The computing device (10) of claim 1, wherein an analysis result of the excluded medical image from the reading worklist is provided as a report in different form than the worklist.
3. The computing device (10) of claim 1, wherein the medical image classified into not strong normal is added to the reading worklist.
4. The computing device (10) of claim 1, wherein the processor (11) is configured to display an analysis result of the medical image in response to a selection of the medical image in the reading worklist.
5. The computing device (10) of claim 4, wherein the analysis result is provided as secondary capture.
6. The computing device (10) of claim 4 or 5, wherein the analysis result comprises a heatmap visually indicating a position or predicted value of a lesion.
7. A method implemented by a computing device (10) operated by at least one processor (11), the method comprising: obtaining disease prediction scores for different diseases predicted in a medical image using abnormality prediction model (100); determining abnormality scores for the different diseases using the disease prediction scores for the different diseases; calculating cut-off scores for the different diseases which makes a specific reading sensitivity; classifying the medical image into strong normal or not strong normal using the abnormality score; and adding the medical image to a reading worklist depending on the classification result, wherein classifying the medical image comprises: classifying the medical image into the strong normal when the abnormality scores for the different diseases are less than or equal to the cut-off scores corresponding to the different diseases, for all of the different diseases, wherein the medical image classified into strong normal is excluded from the reading working list.
8. The method of claim 7, wherein the analysis result of the excluded medical image from the reading worklist is provided as a report in different form than the reading worklist, or wherein the medical image classified into not strong normal is added to the reading worklist.
9. The method of claim 7, further comprising displaying analysis result of the medical image in response to a selection of the medical image in the reading worklist, wherein the analysis result is provided as secondary capture, and the analysis result comprises a heatmap visually indicating a position or predicted value of a lesion.

Description

BACKGROUND (a) Field The present disclosure relates to methods of classifying a medical and a computing device implementing the methods. (b) Description of the Related Art In a medical field, various products utilizing an artificial intelligence (AI) technology has been developing, and a diagnosis assistance system implemented with the AI-based medical image reading technology is a representative example thereof. The AI-based medical image reading technology can analyze the entire medical image with an AI algorithm and provide an abnormal lesion visually. A specialized doctor for image reading (hereinafter, referred to as a "reader") can be provided with an analysis result of the medical image from the diagnosis assistance system and read the medical image with reference thereto. The reader can check a reading result provided by the diagnosis assistance system and medical records of a patient in a worklist, and can change the image reading order by way of worklist sorting based on specific criteria (e.g., emergency, abnormality, etc.). Using the function of worklist sorting, the reader can preferentially read an image required to be read urgently or an image where an abnormality is detected, rather than an image analyzed as normal. However, since the function worklist sorting is only to change the reading order of the images already included in the worklist, reading a normal image should be done in the end. Therefore, workload of the reader does not change. In addition, although the reading level of the diagnosis assistance system has been increasing, reading sensitivity of the diagnosis assistance system is not very high due to the trade-off between the sensitivity and specificity. The US 2020/160983 A1 discloses a medical scan triaging system operable to generate a global abnormality probability for each of a plurality of medical scans by utilizing a computer vision model trained on a training set of medical scans. A triage probability threshold is determined based on user input to a client device. A first subset of the plurality of medical scans, designated for human review, is determined by identifying medical scans with a corresponding global abnormality probability that compares favorably to the triage probability threshold. A second subset of the plurality of medical scans, designated as normal, is determined by identifying ones of the plurality of medical scans with a corresponding global abnormality probability that compares unfavorably to the triage probability threshold. SUMMARY The problem of the present invention is solved by a computing device according to independent claim 1 as well as by a method implemented by a computing device according to independent claim 7. The dependent claims refer to further advantageous developments of the present invention. The method may further include adding the analysis result to a reading worklist, and the input image whose abnormality score is less than or equal to the cut-off score may be not added to the reading worklist. In a case of the input image whose abnormality score is less than or equal to the cutoff score, a filtering result may be added to a separate report from the reading worklist. Obtaining the abnormality score may include, when obtaining disease prediction scores for different diseases from the abnormality prediction model, aggregating the disease prediction scores to determine the abnormality score. Obtaining the abnormality score may include, obtaining calibrated disease prediction scores based on calibration that converts a cut-off score for each disease, which makes the specific reading sensitivity, into the cut-off score, and determining a maximum value among the calibrated disease prediction scores as the abnormality score. In a case where each of disease prediction scores for different diseases is obtained as an abnormality score for each disease from the abnormality prediction model, filtering the input image may include calculating a cut-off score for each disease which makes the specific reading sensitivity, and filtering the input image when the abnormality score for each disease is less than or equal to the cut-off score for a corresponding disease, for all of the different diseases. The abnormality prediction model may include a feature extraction model trained to output a feature of the input image, and at least one disease prediction head model trained to predict at least one disease based on features output from the feature extraction model. The abnormality prediction model may have a sensitivity between 90 % and 100 %. The method may further include obtaining an analysis result of the input image using a classification model that distinguishes between weak normal and abnormal when the abnormality score of the input image is greater than the cut-off score. The method may further include adding the analysis result to a reading worklist, and the input image classified into strong normal may be not added to the reading worklist