Performance estimation of machine learning models that analyze medical images
There is provided a method, comprising: feeding target metadata parameters of target medical image(s) into a performance estimation machine learning (ML) model, and obtaining a performance metric of a computer vision ML model as an outcome of the performance estimation ML model, wherein the performance estimation ML model is trained on a training dataset comprising records, wherein a record is created by feeding a sample medical image into the computer vision ML model, obtaining an outcome of the computer vision ML model, obtaining a ground truth of the sample medical image corresponding to the outcome, extracting metadata parameters associated with the sample medical image, and wherein the record includes metadata parameters associated with the sample medical image and a ground truth label comprising a performance metric of the computer vision ML model computed based on the ground truth of the sample medical image and the outcome.
1 . A computer implemented method of predicting performance of a computer vision machine learning (ML) model, comprising:
feeding a plurality of target metadata parameters of at least one target medical image into a performance estimation ML model; and
obtaining a performance metric of a computer vision ML model as an outcome of the performance estimation ML model,
wherein the performance estimation ML model is trained on a training dataset comprising a plurality of records, wherein a record is created by:
feeding a sample medical image into the computer vision ML model,
obtaining an outcome of the computer vision ML model,
obtaining a ground truth of the sample medical image corresponding to the outcome,
extracting a plurality of metadata parameters associated with the sample medical image, and
wherein the record includes a plurality of metadata parameters associated with the sample medical image and a ground truth label comprising a performance metric of the computer vision ML model computed based on the ground truth of the sample medical image and the outcome.
2 . The computer implemented method of claim 1 , wherein the at least one target medical image is not fed into the performance estimation ML model, the training dataset excludes medical images, and the at least one target medical image is not fed into the computer vision ML model for predicting the performance of the computer vision ML model.
3 . The computer implemented method of claim 1 , wherein the at least one target medical image associated with the plurality of target metadata parameters have not yet been captured during computation of the performance metric, and the at least one target medical image is captured after computation of the performance metric.
4 . The computer implemented method of claim 1 , wherein the at least one target medical image and the sample medical image depict an interior anatomy of subjects, and are captured by a medical image device, wherein the outcome and the ground truth of the sample medical image are a visual medical finding.
5 . The computer implemented method of claim 4 , wherein the metadata parameters are physical attributes of the subjects.
6 . The computer implemented method of claim 4 , wherein the metadata parameters are parameters of at least one of: properties of slices of a 3D medical image, a protocol of capturing the medical images, and physical properties of the medical imaging device.
7 . The computer implemented method of claim 1 , wherein the metadata parameters are selected from a group comprising: fields defined by DICOM® standard, manufacturer of medical imaging device that captured the medical image, model of medical imaging device that captured the medical image, filter type, sex of subject depicted in medical image, age of subject, race of subject, position of subject during acquisition of medical image, whether contrast was administered, and peak kilovoltage (kVp) for x-ray images.
8 . The computer implemented method of claim 1 , wherein the metadata parameters are non-extractable from the medical image.
9 . The computer implemented method of claim 1 , wherein the at least one target medical image comprises a plurality of medical images with different combinations of metadata parameters, wherein feeding and obtaining comprise feeding each combination of metadata parameters to obtain a respective intermediate performance metric, and further comprising computing an average of the intermediate performance metrics.
10 . The computer implemented method of claim 1 , wherein the computer vision ML model is of a plurality of candidate computer vision ML models each associated with a respective corresponding performance estimation ML model, wherein the feeding and the obtaining is performed for each respective corresponding performance estimation ML model of the plurality of candidate computer vision ML models, and further comprising selecting a certain candidate computer vision ML model having a highest performance metric outcome of the respective corresponding performance estimation ML model.
11 . The computer implemented method of claim 1 , further comprising a plurality of performance estimation ML models each trained for predicting a different performance metric for the computer vision ML model, and further comprising obtaining an input from a user defining a tradeoff between the plurality of performance estimation ML models.
12 . The computer implemented method of claim 11 , wherein the plurality of performance estimation ML models comprises a first performance estimation ML model for predicting sensitivity and a second performance estimation ML model for predicting specificity, and further comprising providing a graphical user interface (GUI) configured for enabling the user to select the tradeoff between sensitivity and specificity.
13 . The computer implemented method of claim 1 , further comprising computing an optimal set of metadata parameters, and selecting medical images satisfying the set for feeding into the computer vision ML model.
14 . The computer implemented method of claim 13 , wherein the optimal set is found by iteratively feeding different combinations of metadata parameters into the performance estimation ML model and identifying the optimal set as a certain combination of metadata parameters having highest performance metric.
15 . The computer implemented method of claim 13 , further comprising applying a machine learning interpretability process to the performance estimation ML model for identifying at least one most significant metadata parameter that most contributes to the performance metric outcome, and computing values for the at least one most significant metadata parameter that generate the highest performance metric outcome.
16 . The computer implemented method of claim 13 , wherein the optimal set of metadata parameters includes at least one image acquisition parameter indicating at least one of: image capture protocol, image storage parameter, and an image capture parameter of the medical imaging device, and further comprising generating instructions for adjusting the at least one image acquisition parameter for obtaining the optimal set of metadata parameters in additional images.
17 . The computer implemented method of claim 1 , wherein the metadata parameters and/or records are represented as tabular data, and the performance estimation ML model is designed for processing tabular data.
18 . The computer implemented method of claim 1 , wherein the performance metric comprises sensitivity.
19 . The computer implemented method of claim 18 , wherein the training dataset is created by:
selecting medical images having ground truth of the sample medical image indicating a positive finding depicted in the sample medical image;
setting the performance metric to a value indicating highest performance when the outcome of the computer vision ML model correct identifies the positive finding, or setting the performance metric to a value indicating lowest performance when the outcome of the computer vision ML model fails to correctly identify the positive finding, wherein an average of the performance metric indicates sensitivity of the computer visional ML model for a training benchmark.
20 . The computer implemented method of claim 1 , wherein the performance metric comprises specificity.
21 . The computer implemented method of claim 20 , wherein the training dataset is created by:
selecting medical images having ground truth of the sample medical image indicating lack of a positive finding depicted in the sample medical image;
setting the performance metric to a value indicating highest performance when the outcome of the computer vision ML model correct identifies the lack of the positive finding, or setting the performance metric to a value indicating lowest performance when the outcome of the computer vision ML model fails to correctly identifies the lack of positive finding.
22 . A computer implemented method of training a performance estimation ML model for estimating performance of a computer vision ML model, comprising:
for each medical image of a plurality of medical images:
feeding the medical image into the computer vision ML model,
obtaining an outcome of the computer vision ML model,
obtaining a ground truth of the medical image corresponding to the outcome, and
extracting a plurality of metadata parameters associated with the medical image;
creating a training dataset comprising a plurality of records, wherein a record of an medical image of the plurality of medical images includes the plurality of metadata parameters and a ground truth label comprising a performance metric of the computer vision ML model computed based on the ground truth of the medical image and the outcome; and
training the performance estimation ML model on the training dataset for estimating a performance of the computer vision ML model for target medical images in response to an input of a plurality of target metadata parameters.
23 . A system for predicting performance of a computer vision machine learning (ML) model, comprising:
at least one processor executing a code for:
feeding a plurality of target metadata parameters of at least one target medical image into a performance estimation ML model; and
obtaining a performance metric of a computer vision ML model as an outcome of the performance estimation ML model,
wherein the performance estimation ML model is trained on a training dataset comprising a plurality of records, wherein a record is created by
feeding a sample medical image into the computer vision ML model,
obtaining an outcome of the computer vision ML model,
obtaining a ground truth of the sample medical image corresponding to the outcome,
extracting a plurality of metadata parameters associated with the sample medical image, and
wherein the record includes a plurality of metadata parameters associated with the sample medical image and a ground truth label comprising a performance metric of the computer vision ML model computed based on the ground truth of the sample medical image and the outcome.