Re-weighted self-influence for labeling noise removal in medical imaging data
Described are techniques for image processing. For instance, a process can include obtaining a plurality of labeled input images and determining a threshold percentage associated with the plurality of labeled input images, indicative of a percentage of correctly labeled input images. The process can include determining a respective self-influence for each respective labeled input image included in the plurality of input images and generating a respective self-influence weight for each respective labeled input image, based on the respective self-influence and the threshold percentage associated with each respective labeled input image. The process can include determining one or more loss values using a loss function associated with training a machine learning network based on using the plurality of labeled input images as a training data set, wherein the loss function determines the one or more loss values based on weighting each respective labeled input image by its respective self-influence weight.
1 . An apparatus for processing image data, the apparatus comprising:
at least one memory; and
at least one processor coupled to the at least one memory, the at least one processor configured to:
obtain a plurality of labeled input images comprising medical imaging data labeled with pathology detection grades selected from a set of pathology severity classifications associated with a detected pathology;
determine self-influence threshold information corresponding to the plurality of labeled input images and a pathology detection machine learning network trained on the plurality of labeled input images, wherein the self-influence threshold information includes a corresponding self-influence threshold level for each classification of the set of pathology severity classifications;
determine a respective self-influence for each respective labeled input image included in the plurality of input images, wherein the respective self-influence is determined using the respective labeled input image as a training sample and as a test sample to track loss reduction information between corresponding loss gradients across a plurality of training checkpoints configured for a baseline training procedure of the pathology detection machine learning network;
generate a respective self-influence weight for each respective labeled input image, based on a piecewise weighting according to the respective self-influence determined from the loss reduction information and the corresponding self-influence threshold level for the pathology severity classification associated with each respective labeled input image; and
re-train the pathology detection machine learning network to update model parameters and generate a re-trained pathology detection machine learning network from the updated model parameters, the retraining based on a self-influence loss function and a ranked re-weighted training data set comprising each respective labeled input image weighted by its respective self-influence weight.
2 . The apparatus of claim 1 , wherein, to generate the respective self-influence weight for each respective labeled input image, the at least one processor is configured to:
generate the respective self-influence weight using a value of one based on a determination that the respective self-influence is less than the corresponding self-influence threshold level for the pathology severity classification associated with the respective labeled input image; and
generate the respective self-influence weight using a first ratio of a self-influence associated with the corresponding self-influence threshold level and the respective self-influence, based on a determination that the respective self-influence is greater than the corresponding self-influence threshold level.
3 . The apparatus of claim 2 , wherein the at least one processor is further configured to:
generate the respective self-influence weight using a second ratio of the self-influence associated with the corresponding self-influence threshold level and the respective self-influence, based on a determination that the first ratio is greater than the second ratio.
4 . The apparatus of claim 3 , wherein the respective self-influence weight generated using the first ratio is a first reduced weighting self-influence weight and the respective self-influence weight generated using the second ratio is a second reduced weighting self-influence weight, the second reduced weighting self-influence weight less than the first reduced weighting self-influence weight.
5 . The apparatus of claim 1 , wherein the corresponding self-influence threshold level for a pathology severity classification is determined based on a noise level associated with a portion of the plurality of labeled input images having the pathology severity classification, and wherein the corresponding self-influence threshold level is indicative of a percentage of the portion of the plurality of labeled input images having the pathology severity classification that do not include labeling noise.
6 . The apparatus of claim 1 , wherein:
the plurality of labeled input images comprises a plurality of labeled medical images; and
each respective labeled medical image is associated with a respective labeled class of a plurality of labeled classes included in the plurality of labeled medical images.
7 . The apparatus of claim 6 , wherein the at least one processor is configured to:
determine a respective set of self-influence values and a respective set of corresponding self-influence weights for labeled input images associated with each respective labeled class of the plurality of labeled classes.
8 . The apparatus of claim 6 , wherein each respective labeled class of the plurality of labeled medical images is associated with a different corresponding self-influence threshold level and a different range of respective self-influence values.
9 . A method for processing image data, the method comprising:
obtaining a plurality of labeled input images comprising medical imaging data labeled with pathology detection grades selected from a set of pathology severity classifications associated with a detected pathology;
determining self-influence threshold information corresponding to the plurality of labeled input images and a pathology detection machine learning network trained on the plurality of labeled input images, wherein the self-influence threshold information includes a corresponding self-influence threshold level for each classification of the set of pathology severity classifications;
determining a respective self-influence for each respective labeled input image included in the plurality of input images, wherein the respective self-influence is determined using the respective labeled input image as a training sample and as a test sample to track loss reduction information between corresponding loss gradients across a plurality of training checkpoints configured for a baseline training procedure of the pathology detection machine learning network;
generating a respective self-influence weight for each respective labeled input image, based on a piecewise weighting according to the respective self-influence determined from the loss reduction information and the corresponding self-influence threshold level for the pathology severity classification associated with each respective labeled input image; and
re-training the pathology detection machine learning network to update model parameters and generate a re-trained pathology detection machine learning network from the updated model parameters, the retraining based on a self-influence loss function and a ranked re-weighted training data set comprising each respective labeled input image weighted by its respective self-influence weight.
10 . The method of claim 9 , wherein generating the respective self-influence weight for each respective labeled input image comprises:
generating the respective self-influence weight using a value of one based on a determination that the respective self-influence is less than the corresponding self-influence threshold level for the pathology severity classification associated with the respective labeled input image; and
generating the respective self-influence weight using a first ratio of a self-influence associated with the corresponding self-influence threshold level and the respective self-influence, based on a determination that the respective self-influence is greater than the corresponding self-influence threshold level.
11 . The method of claim 10 , further comprising:
generating the respective self-influence weight using a second ratio of the self-influence associated with the corresponding self-influence threshold level and the respective self-influence, based on a determination that the first ratio is greater than the second ratio.
12 . The method of claim 11 , wherein the respective self-influence weight generated using the first ratio is a first reduced weighting self-influence weight and the respective self-influence weight generated using the second ratio is a second reduced weighting self-influence weight, the second reduced weighting self-influence weight less than the first reduced weighting self-influence weight.
13 . The method of claim 9 , wherein the corresponding self-influence threshold level for a pathology severity classification is determined based on a noise level associated with a portion of the plurality of labeled input images having the pathology severity classification, and wherein the corresponding self-influence threshold level is indicative of a percentage of the portion of the plurality of labeled input images having the pathology severity classification that do not include labeling noise.
14 . The method of claim 9 , wherein:
the plurality of labeled input images comprises a plurality of labeled medical images; and
each respective labeled medical image is associated with a respective labeled class of a plurality of labeled classes included in the plurality of labeled medical images.
15 . The method of claim 14 , further comprising:
determining a respective set of self-influence values and a respective set of corresponding self-influence weights for labeled input images associated with each respective labeled class of the plurality of labeled classes.
16 . The method of claim 14 , wherein each respective labeled class of the plurality of labeled medical images is associated with a different threshold percentage and a different range of respective self-influence values.
17 . A non-transitory computer-readable storage medium having stored therein instructions which, when executed by a processor, cause the processor to perform operations comprising:
obtaining a plurality of labeled input images comprising medical imaging data labeled with pathology detection grades selected from a set of pathology severity classifications associated with a detected pathology;
determining self-influence threshold information corresponding to the plurality of labeled input images and a pathology detection machine learning network trained on the plurality of labeled input images, wherein the self-influence threshold information includes a corresponding self-influence threshold level for each classification of the set of pathology severity classifications;
determining a respective self-influence for each respective labeled input image included in the plurality of input images, wherein the respective self-influence is determined using the respective labeled input image as a training sample and as a test sample to track loss reduction information between corresponding loss gradients across a plurality of training checkpoints configured for a baseline training procedure of the pathology detection machine learning network;
generating a respective self-influence weight for each respective labeled input image, based on a piecewise weighting according to the respective self-influence determined from the loss reduction information and the corresponding self-influence threshold level for the pathology severity classification associated with each respective labeled input image; and
re-training the pathology detection machine learning network to update model parameters and generate a re-trained pathology detection machine learning network from the updated model parameters, the retraining based on a self-influence loss function and a ranked re-weighted training data set comprising each respective labeled input image weighted by its respective self-influence weight.
18 . The non-transitory computer-readable storage medium of claim 17 , wherein generating the respective self-influence weight for each respective labeled input image comprises:
generating the respective self-influence weight using a value of one based on a determination that the respective self-influence is less than the corresponding self-influence threshold level for the pathology severity classification associated with the respective labeled input image; and
generating the respective self-influence weight using a first ratio of a self-influence associated with the corresponding self-influence threshold level and the respective self-influence, based on a determination that the respective self-influence is greater than the corresponding self-influence threshold level.
19 . The non-transitory computer-readable storage medium of claim 17 , wherein the corresponding self-influence threshold level for a pathology severity classification is determined based on a noise level associated with a portion of the plurality of labeled input images having the pathology severity classification, and wherein the corresponding self-influence threshold level is indicative of a percentage of the portion of the plurality of labeled input images having the pathology severity classification that do not include labeling noise.
20 . The non-transitory computer-readable storage medium of claim 17 , wherein:
the plurality of labeled input images comprises a plurality of labeled medical images;
each respective labeled medical image is associated with a respective labeled class of a plurality of labeled classes included in the plurality of labeled medical images; and
the processor is further configured to perform operations comprising determining a respective set of self-influence values and a respective set of corresponding self-influence weights for labeled input images associated with each respective labeled class of the plurality of labeled classes.