Medical data collection for machine learning
A system ( 100 ) and computer-implemented method are provided for data collection for distributed machine learning of a machine learnable model. A privacy policy data ( 050 ) is provided defining computer-readable criteria for limiting a selection of medical image data ( 030 ) to a subset of the medical image data to obfuscate an identity of the at least one patient. The medical image data is selected based on the computer-readable criteria to obtain privacy policy-compliant training data ( 060 ) for transmission to another entity. The system and method enable medical data collection at clinical sites without requiring manual oversight, and enables such selections to be made automatically, e.g., based on a request for medical image data which may be received from outside of the clinical site.
1. A system for data collection for machine learning of a machine learnable model, the system comprising:
one or more hardware processors;
a communication interface to another entity associated with training of a machine learnable model; and
a memory storing instructions to:
receive a request for training data from the other entity;
access:
medical image data of at least one patient and label data defining labels associated with the medical image data; and
privacy policy data defining one or more computer-readable criteria for limiting a selection of the medical image data to a subset of the medical image data to obfuscate an identity of the at least one patient;
verify whether the request is in compliance with the privacy policy data;
based on the one or more computer-readable criteria, automatically perform, by the one or more hardware processors, the selection of the medical image data and an associated selection of the label data to obtain privacy policy-compliant training data, wherein the selection is limited to one or more image regions of the medical image data; and
via the communication interface, transmit the privacy policy-compliant training data to the other entity to enable the machine learnable model to be trained on the basis of the privacy policy-compliant training data.
2. The system according to claim 1 , wherein the one or more computer-readable criteria limit at least one of the group of:
a number,
an individual or aggregate size, and
a distribution,
of the one or more image regions per image or per patient.
3. The system according to claim 1 , wherein the memory stores further instructions to apply a machine learning data augmentation technique to the medical image data before selecting and transmitting the privacy policy-compliant training data to the other entity.
4. The system according to claim 1 , wherein the memory stores further instructions to, before or when generating the privacy policy-compliant training data, obfuscate one or more of whether:
different subsets of the medical data, or different parts of a subset, belong to a same patient; or
the subset of the medical image data, or a part of the subset, is augmented by a machine learning data augmentation technique.
5. The system according to claim 4 , wherein the memory stores further instructions to randomize or pseudo-randomize an identifier of each of the different subsets of the medical image data, and/or an identifier of each of the different parts of the subset.
6. The system according to claim 1 , wherein the memory stores further instructions to determine whether the request, individually and/or in aggregate with previous requests, exceeds said limitation defined by the one or more computer-readable criteria, and when the limitation is exceeded, reject the request.
7. The system according to claim 1 , wherein the memory stores further instructions to:
via the network interface, receive a processing algorithm for processing the selection of the medical image data;
execute the processing algorithm to obtain a processing result; and
via the network interface, transmit the processing result as training data to a training system which is configured to train the machine learnable model based on the processing result provided by the training data.
8. A computer-implemented method for data collection for distributed machine learning of a machine learnable model, the method comprising:
receiving, by a system comprising one or more processors, a request for training data;
accessing:
medical image data of at least one patient and label data defining labels associated with the medical image data;
privacy policy data defining one or more computer-readable criteria for limiting a selection of the medical image data to a subset of the medical image data to obfuscate an identity of the at least one patient;
verifying whether the request is in compliance with the privacy policy data;
based on the one or more computer-readable criteria, automatically performing, by the one or more processors, the selection of the medical image data and an associated selection of the label data to obtain privacy policy-compliant training data, wherein the selection is limited to one or more image regions of the medical image data; and
via a communication interface, transmitting the privacy policy-compliant training data to another entity to enable the machine learnable model to be trained on the basis of the privacy policy-compliant training data.
9. A non-transitory computer-readable storage medium comprising instructions arranged to cause one or more processors to:
receive a request for training data;
access medical image data of at least one patient and label data defining labels associated with the medical image data;
access privacy policy data defining one or more computer-readable criteria for limiting a selection of the medical image data to a subset of the medical image data to obfuscate an identity of the at least one patient;
verify whether the request is in compliance with the privacy policy data;
based on the one or more computer-readable criteria, automatically perform the selection of the medical image data and an associated selection of the label data to obtain privacy policy-compliant training data, wherein the selection is limited to one or more image regions of the medical image data; and
via a communication interface, transmit the privacy policy-compliant training data to another entity to enable the machine learnable model to be trained on the basis of the privacy policy-compliant training data.