Active machine learning for training an event classification
An event classification is trained by machine learning. An anomaly detection for detecting events in an image data set is thereby performed. Based on the performance of the anomaly detection, a model assumption of the event classification is determined. An image data set may include a plurality of images, and each image may include an array of pixels. Further, an image data set may include volume data and/or a time sequence of images and in this way represent a video sequence.
1. A method for training an event classification by machine learning, the method comprising:
determining a model assumption of the event classification by iteratively adapting a candidate subset during a supervised learning step, wherein the determining of the model assumption of the event classification comprises:
(a) performing an anomaly detection based on the model assumption to detect events in an image data set;
(b) determining the candidate subset with the anomaly detection;
(c) selecting a current image of the candidate subset;
(d) performing the supervised learning step based on the current image of the candidate subset of the image data set with a user interface;
(e) adapting the model assumption based on the supervised learning step; and
(f) repeating steps (a) to (e) according to the adapted model assumption of step (e).
2. The method as claimed in claim 1 , further comprising:
determining the candidate subset of the image data set based on a difference between a mean value of a prediction of the anomaly detection for image features of a specific image and the image features of the specific image.
3. The method as claimed in claim 1 , further comprising:
determining the candidate subset of the image data set based on a Kullback-Leibler divergence between a prediction of the anomaly detection for image features of a specific image and the image features of the specific image.
4. The method as claimed in claim 1 , further comprising:
determining the candidate subset of the image data set based on a method of drawing lots, which assigns a probability to each image of the image data set.
5. The method as claimed in claim 4 , further comprising:
for each image of the image data set, determining the corresponding assigned probability of the method of drawing lots based on at least one element that is selected from the group consisting of:
a difference between a mean value of a prediction of the anomaly detection for image features of the corresponding image and the image features of the corresponding image, and
a Kullback-Leibler divergence between a prediction of the anomaly detection for image features of the corresponding image and the image features of the corresponding image.
6. The method as claimed in claim 1 , further comprising:
statistically determining the candidate subset before the supervised learning step.
7. The method as claimed in claim 1 , wherein the performing of the supervised learning step comprises:
sequentially outputting the images of the candidate subset of the image data set by the user interface; and
for each image of the candidate subset of the image data set, inquiring by the user interface whether there is an event in the corresponding image.
8. The method as claimed in claim 7 , wherein the sequential outputting of the images of the candidate subset of the image data set comprises:
for at least some of the images of the candidate subset of the image data set, marking an event candidate in the output of the respective image.
9. The method as claimed in claim 7 , wherein the inquiry as to whether an event is present is a yes/no inquiry.
10. The method as claimed in claim 7 , wherein the sequential outputting of the images of the candidate subset of the image data set is performed in a random sequence.
11. The method as claimed in claim 1 , further comprising:
performing an unsupervised or weakly supervised learning step based on a training image data set for determining a model assumption of the anomaly detection.
12. The method as claimed in claim 1 , further comprising:
performing the event classification for classifying events in a further image data set based on the determined model assumption of the event classification,
wherein the image data set and the further image data set at least partially overlap or are disjoint.
13. The method as claimed in claim 1 , wherein the events are selected from the group consisting of: a mitosis of a cell, an apoptosis of a cell, and a defect of a test piece.
14. The method as claimed in claim 1 , further comprising:
obtaining the image data set from a detector selected from the group consisting of: a computed tomography scanner, an optical phase-contrast microscope, an optical microscope, an optical microscope with contrast imaging, an optical microscope with structured illumination, a 3D detector, a time-of-flight detector, a stereoscopic detector, a phase-shifting detector, and an interferometer.
15. The method as claimed in claim 1 , wherein the current image of the candidate subset is selected by drawing lots.
16. A method for training an event classification by machine learning, the method comprising:
performing an anomaly detection for detecting events in an image data set;
determining a model assumption of the event classification based on the performing of the anomaly detection, wherein the determining of the model assumption of the event classification comprises performing a supervised learning step based on a candidate subset of the image data set by a user interface, and wherein the performing of the supervised learning step comprises (a) sequentially outputting the images of the candidate subset of the image data set by the user interface, and (b) for each image of the candidate subset of the image data set, inquiring by the user interface whether there is an event in the corresponding image;
iteratively adapting the candidate subset during the supervised learning step; and
determining a sequence with which the sequential outputting of the images of the candidate subset of the image data set takes place based on elements that are selected from the group consisting of:
a difference between a mean value of a prediction of the anomaly detection for image features of a specific image and the image features of the specific image; and
a Kullback-Leibler divergence between a prediction of the anomaly detection for pixel values of a specific image and the image features of the specific image.
17. An apparatus for training an event classification by machine learning, the apparatus comprising:
a memory configured to store an image data set;
a processor configured to:
determine a model assumption of the event classification by iteratively adapting a candidate subset during a supervised learning step, wherein, to determine the model assumption of the event classification, the processor is configured to:
(a) perform an anomaly detection based on the model assumption to detect events in an image data set;
(b) determine the candidate subset with the anomaly detection;
(c) select a current image of the candidate subset;
(d) perform the supervised learning step based on the current image of the candidate subset of the image data set with a user interface;
(e) adapt the model assumption based on the supervised learning step; and
(f) repeat steps (a) to (e) according to the adapted model assumption of step (e).
18. The apparatus as claimed in claim 17 , further comprising a phase-contrast microscope.
19. The apparatus as claimed in claim 17 , wherein the current image of the candidate subset is selected by drawing lots.