Self-supervised learning for medical image quality control
Provided herein are methods for automated image quality control (QC). The method comprises: generating training data based at least in part on metadata obtained from a data augmentation process; training a model for a QC task based at least in part on the training data. The model is trained using a self-supervised learning algorithm.
1 . A computer-implemented method for automated image quality control (QC), the method comprising:
generating training data based at least in part on metadata obtained from a data augmentation process; and
training a model for a QC task based at least in part on the training data, wherein the QC task comprises an out-of-distribution (OOD) detection,
wherein the model is trained using a self-supervised learning algorithm and is trained to predict metadata from an input image and
i) wherein an OOD event is determined when the predicted metadata does not match the metadata or
ii) wherein an OOD event is determined when an uncertainty score of the metadata prediction is beyond a predetermined threshold.
2 . The computer-implemented method of claim 1 , wherein supervised learning algorithm is contrastive learning.
3 . The computer-implemented method of claim 1 , wherein the metadata is extracted from a header of an image.
4 . The computer-implemented method of claim 3 , wherein the metadata is used to generate a label for the image.
5 . The computer-implemented method of claim 1 , wherein the QC task is image registration quality control.
6 . The computer-implemented method of claim 5 , wherein the model is trained to predict an alignment based on a similarity map between embeddings of two input images.
7 . The computer-implemented method of claim 6 , wherein the embeddings are local embeddings or global embeddings produced by an encoder-decoder network.
8 . A computer-implemented method for automated image quality control (QC), the method comprising:
generating training data based at least in part on metadata obtained from a data augmentation process, wherein the data augmentation process comprises generating cropped patch with simulated artifact; and
training a model for a QC task based at least in part on the training data, wherein the model is trained using a self-supervised learning algorithm.
9 . The computer-implemented method of claim 8 , further comprising replacing a corresponding patch in an input image with the cropped patch with the simulated artifact.
10 . A computer-implemented method for automated image quality control (QC), the method comprising:
generating training data based at least in part on metadata obtained from a data augmentation process, wherein the training data comprises 3D image including a stack of slices and wherein generating the training data comprises generating a label for the 3D image using a multiple-instance-learning method; and
training a model for a QC task based at least in part on the training data, wherein the model is trained using a self-supervised learning algorithm.
11 . The computer-implemented method of claim 10 , further comprising determining a pooling strategy for combining embeddings of one or more slices based on a selected multiple-instance-learning assumption.