Trained model creation method using machine learning used to perform a segmentation process on a cell image and trained model creation system using machine learning used to perform a segmentation process on a cell image
A trained model creation method includes creating a trained model ( 50 ) based on training data ( 86 ) including first teacher image data ( 86 b ), determining whether or not an estimated region is erroneously estimated in inference result data ( 84 ), determining a noise region ( 60 ) included in the first teacher image data ( 86 b ), updating the first teacher image data ( 86 b ), and re-creating trained model ( 55 ).
1 . A trained model creation method using machine learning used to perform a segmentation process on a cell image, the trained model creation method comprising:
creating a trained model based on training data including first input image data and first teacher image data that corresponds to the first input image data and is labeled with at least a first label indicating a detection target and a second label indicating a background;
executing an inference process of inputting second input image data into the created trained model and outputting inference result data;
determining whether or not a region estimated to be the detection target in the inference result data output by the inference process is erroneously estimated, based on the inference result data and second teacher image data that corresponds to the second input image data and is labeled with at least the first label and the second label;
specifying one or more first noise regions in which the estimated region is determined to be erroneously estimated;
determining one or more second noise regions corresponding to the one or more first noise regions in the first teacher image data;
further labeling the one or more second noise regions in the first teacher image data with a third label, which is a different type from the first label and the second label, indicating a noise region and updating the first teacher image data; and
re-creating a trained model based on updated training data including the updated first teacher image data and the first input image data.
2 . The trained model creation method according to claim 1 , further comprising:
creating, as image data for extracting the region erroneously estimated in the inference result data, subtraction image data indicating a difference between the second teacher image data and the inference result data corresponding to the second teacher image data after the executing the inference process and before the determining whether or not the region is erroneously estimated.
3 . The trained model creation method according to claim 2 , wherein the creating the subtraction image data includes creating the subtraction image data in which a first region labeled with the first label in both the second teacher image data and the inference result data among corresponding regions of the second teacher image data and the inference result data, a second region labeled with the second label in the second teacher image data and labeled with the first label in the inference result data among the corresponding regions of the second teacher image data and the inference result data, and a third region labeled with the first label in the second teacher image data and labeled with the second label in the inference result data among the corresponding regions of the second teacher image data and the inference result data identifiably appear.
4 . The trained model creation method according to claim 1 , wherein the one or more first noise regions are not all of regions erroneously estimated.
5 . The trained model creation method according to claim 1 , wherein the re-creating the trained model includes re-creating the trained model based on the updated training data to which, in addition to the updated training data, training data including the first input image data and the first teacher image data labeled with the third label has been further added.
6 . The trained model creation method according to claim 1 , further comprising:
displaying, on a display, the region erroneously estimated in the inference result data after the executing the inference process and before the determining whether or not the region is erroneously estimated.
7 . The trained model creation method according to claim 6 , wherein the displaying, on the display, the region erroneously estimated includes displaying, on the display, the region erroneously estimated and satisfying a preset condition regarding at least one of an aspect ratio, an exclusive area, or shape complexity.
8 . The trained model creation method according to claim 1 , wherein the determining whether or not the region is erroneously estimated includes determining whether or not a region estimated to be a low-occurrence detection target having an occurrence frequency equal to or less than a predetermined threshold and included in the detection target in the inference result data is erroneously estimated.
9 . The trained model creation method according to claim 8 , wherein the low-occurrence detection target includes an abnormal cell.
10 . The trained model creation method according to claim 9 , further comprising:
acquiring an occurrence frequency of the detection target included in the training data before the creating the trained model; and
notifying a user that the low-occurrence detection target is included when the low-occurrence detection target is included in the detection target based on an acquired occurrence frequency of the detection target.
11 . The trained model creation method according to claim 10 , wherein the creating the trained model includes creating the trained model based on the training data to which the training data including the low-occurrence detection target has been added.
12 . A trained model creation system using machine learning used to perform a segmentation process on a cell image, the trained model creation system comprising:
a storage to store training data including first input image data and first teacher image data that corresponds to the first input image data and is labeled with at least a first label indicating a detection target and a second label indicating a background, second input image data and second teacher image data that corresponds to the second input image data and is labeled with the first label and the second label;
a trained model creator to create a trained model based on the training data acquired by the storage; and
an inference process executor to execute an inference process of inputting the second input image data stored in the storage into the created trained model and outputting inference result data; wherein
the storage is operable to further store updated training data including updated first teacher image data in which one or more second noise regions in the first teacher image data are further labeled with a third label indicating a noise region, which is a different type label from the first label and the second label, and the first input image data, the one or more second noise regions corresponding to specified one or more first noise regions, the one or more first noise regions in the inference result data which are determined to be erroneously estimated as detection targets based on the inference result data and the second teacher image data; and
the trained model creator is operable to re-create a trained model based on the updated training data stored in the storage.