Anomaly detection in volumetric medical 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 detection of an anomaly in a volumetric medical image, the method comprising:
for a volumetric medical image comprising a spaced sequence of two-dimensional slice images, encoding each two-dimensional slice image using a two-dimensional regular convolutional neural network (CNN), to generate a sequence of encoded slices; and
applying the sequence of encoded slices to a recurrent neural network (RNN), the RNN yielding an output indicative of the presence or absence of an anomaly in the volumetric medical image.
2. The method of claim 1 , in which the CNN has been trained using an output layer that is omitted during the step of encoding each slice.
3. The method of claim 2 , in which the step of encoding each slice comprises outputting encoded slices from a final dense layer of the CNN.
4. The method of claim 1 , in which the RNN is a Long Short-Term Memory network.
5. The method of claim 1 , in which the RNN output comprises a probability of anomaly presence within the volumetric image.
6. The method of claim 1 , in which the RNN output comprises a probability map indicative of the likelihood that locations within the volumetric image correspond to the anomaly.
7. The method of claim 1 , in which the volumetric image is a head CT and the anomaly is intracranial hemorrhage.
8. The method of claim 1 , in which the volumetric image is a chest CT and the anomaly is presence of a lung nodule.
9. The method of claim 1 , in which the volumetric image is a chest CT and the anomaly is presence of a pulmonary embolism.
10. The method of claim 1 , further comprising:
combining the RNN output with outputs from one or more other neural network models to generate an ensemble output indicative of the presence or absence of the anomaly.
11. The method of claim 10 , in which the one or more other neural network models comprise a 3D CNN.
12. The method of claim 10 , in which the one or more other neural network models comprise a second 2D CNN-RNN model having a different architecture.
13. A computing device for anomaly detection in a volumetric medical image, the computing device comprising:
a digital memory;
at least one processor coupled to the digital memory, the at least one processor configured execute instructions stored in the memory to:
encode each slice of a volumetric medical image comprised of a spaced sequence of two-dimensional slice images, using a two-dimensional regular convolutional neural network (CNN), to generate a sequence of encoded slices; and
apply the sequence of encoded slices to a recurrent neural network (RNN), the RNN yielding an output indicative of the presence or absence of an anomaly in the volumetric medical image.
14. The apparatus of claim 13 , in which the encoded slices are extracted from the output of a final dense layer in the CNN, the final dense layer preceding an omitted output layer used for CNN training.
15. The apparatus of claim 14 , in which the RNN is a Long Short-Term Memory network.
16. The apparatus of claim 15 , in which the RNN output comprises a probability of anomaly presence within the volumetric image.
17. The apparatus of claim 15 , in which the RNN output comprises a probability map indicative of the likelihood that locations within the volumetric image correspond to the anomaly.
18. The apparatus of claim 13 , in which the volumetric image is a head CT and the anomaly is intracranial hemorrhage.
19. The apparatus of claim 13 , in which the volumetric image is a chest CT and the anomaly is presence of a lung nodule.