Anomaly detection in volumetric images using sequential convolutional and recurrent neural networks
Computer-implemented methods and apparatuses for anomaly detection in volumetric images are provided. A two-dimensional convolutional neural network (CNN) is used to encode slices within a volumetric image, such as a CT scan. The CNN may be trained using an output layer that is subsequently omitted during use of the CNN as an encoder. The CNN encoder output is applied to a recurrent neural network (RNN), such as a long short-term memory network. The RNN may output various indications of the presence, probability and/or location of anomalies within the volumetric image.
1. A computer-implemented method for detecting anomalies in volumetric medical images, the method comprising:
training an initial computer-implemented two-dimensional regular convolutional neural network (CNN) using a training set comprising volumetric medical images, each comprising a spaced sequence of two-dimensional slice images taken along a defined axis, and a first set of slice-level labels comprising indications of anomolies; and
training a recurrent neural network (RNN) using an RNN training set comprising encoded volumetric medical images and a corresponding second set of labels comprising indications of anomalies, each encoded volumetric medical image comprising a sequence of spaced two-dimensional slice images taken along a defined axis, in which:
the encoded volumetric medical images each comprising a spaced sequence of two-dimensional slice images taken along a defined axis and encoded by a modified CNN; and
the modified CNN comprises the trained initial CNN having an output layer removed, leaving a last dense layer prior to the removed output layer as a new output layer, such that each encoded slice has the dimension of the new output layer.
2. The method of claim 1 , in which the first set of slice-level labels comprise a binary indicator of whether a slice contains evidence of a disease.
3. The method of claim 1 , in which the first set of slice-level labels comprise a segmentation mask marking evidence of a disease in a slice.
4. The method of claim 1 , in which the RNN is a Long Short-Term Memory RNN.
5. The method of claim 1 , in which the second set of labels comprises a single series-level label.
6. The method of claim 5 , in which the single series-level label comprises an indication of whether a volumetric image contains a given disease.
7. The method of claim 6 , in which the RNN output comprises a probability that a given volumetric image contains the given disease.
8. The method of claim 1 , in which the second set of labels comprises a sequence of slice-level indicators of the presence of a given disease.
9. The method of claim 8 , in which the RNN output comprises slice-by-slice probabilities of the presence of the given disease.
10. The method of claim 8 , in which the RNN output comprises a segmentation map indicative of disease location.
11. A computer-implemented method for detecting anomalies in volumetric medical images, the method comprising: determining initial coefficients for a regular convolutional neural network (CNN) that is part of a neural network sequence comprised of the regular CNN and a recurrent neural network (RNN) where the regular CNN has an output layer feeding an input layer of the RNN, the initial coefficients determined at least in part by training an initial CNN as a feature extractor using 2D medical images; loading initial coefficients into the regular CNN portion of the neural network sequence; and training the neural network sequence as a whole using a training set of volumetric medical images each comprising a spaced sequence of two-dimensional slice images taken along a defined axis, and a corresponding set of training labels indicative of anomalies within the volumetric images.
12. The method of claim 11 , in which the regular CNN comprises an output layer that is a final dense layer.
13. The method of claim 11 , in which the training labels comprise, for each volumetric image, a binary indicator of whether the volumetric image contains evidence of a disease.
14. The method of claim 11 , in which the training labels comprise, for each volumetric image, a segmentation mask marking evidence of a disease in the volumetric image.
15. The method of claim 11 , in which the RNN is a Long Short-Term Memory RNN.
16. The method of claim 11 , further comprising:
applying a volumetric medical image to the neural network sequence input, and generating a resulting output associated with the detection of an anomaly in the volumetric medical image.
17. The method of claim 16 , in which the neural network sequence output comprises a probability that a given volumetric image contains evidence of a given disease.
18. The method of claim 16 , in which the neural network sequence output comprises a segmentation map indicative of disease location.