Multi-modal medical image processing
The present invention relates to the identification of regions of interest in medical images. More particularly, the present invention relates to the identification of regions of interest in medical images based on encoding and/or classification methods trained on multiple types of medical imaging data. Aspects and/or embodiments seek to provide a method for training an encoder and/or classifier based on multimodal data inputs in order to classify regions of interest in medical images based on a single modality of data input source.
1. A method for automatically identifying regions of interest in medical or clinical image data, the method comprising:
receiving unlabelled input data at a multimodal neural encoder comprising a trained encoder and a trained joint representation module, the unlabelled input data comprising data from one or more different modalities of unlabelled input data, wherein the one or more different modalities of unlabelled unlabeled input data is one fewer modality than a number of different modalities of input training data used to train both the trained encoder and the trained joint representation module;
encoding the unlabelled input data using the trained encoder;
determining a joint representation from the encoded unlabelled input data using the trained joint representation module; and
generating labelled data for the unlabelled input data by using the joint representation as an input for a trained classifier, wherein the trained classifier is trained with the input training data,
wherein the trained encoder comprises a different bias for each of the number of different modalities of input training data used to train the trained encoder.
2. The method of claim 1 , wherein the trained encoder comprises a different set of weights for each of the number of different modalities of input training data used to train the trained encoder.
3. The method of claim 1 , wherein only one modality of unlabelled input data is provided to the multimodal neural encoder.
4. The method of claim 1 , wherein the unlabelled input data received by the multimodal neural encoder comprises one or more of: a mammography; an X-ray; a computerised tomography (CT) scan; a magnetic resonance imaging (MRI) data; histology data; mammography data; genetic sequence data; and/or an ultrasound data.
5. The method of claim 1 , wherein the trained joint representation module is trained using one or more outputs received from the trained encoder.
6. The method of claim 1 , wherein the trained joint representation module receives the encoded unlabelled input data as three-dimensional tensors of floating point numbers.
7. The method of claim 1 , wherein the joint representation is in the form of a vector.
8. The method of claim 1 , wherein generating labelled data comprises generating an indication of one or more regions of interest in the unlabelled input data.
9. The method of claim 1 , wherein encoding the unlabelled input data comprises encoding using a plurality of trained encoders.
10. The method of claim 9 , wherein the plurality of trained encoders output a plurality of vectors having a same length used to train the trained joint representation module, each vector of the plurality of vectors representing a different modality of the input training data.
11. The method of claim 1 , wherein the trained encoder of the multimodal neural encoder comprises a convolutional neural network.
12. The method of claim 1 , wherein the input training data used to train the trained joint representation module comprises image data and non-image data.
13. The method of claim 1 , wherein the trained joint representation module of the multimodal neural encoder comprises a learnt pooling operation, and the trained encoder of the multimodal neural encoder comprises a convolutional neural network.
14. A method of classifying data for medical or clinical purposes, comprising:
receiving unlabelled input data at a multimodal neural encoder comprising an encoder and a joint representation module, the unlabelled input data comprising data from one or more different modalities of unlabelled input data, wherein the one or more different modalities of unlabelled input data is one fewer modality than a number of different modalities of input training data used to train both the encoder and the joint representation module;
encoding the unlabelled input data into a joint representation using the encoder and the joint representation module; and
performing classification using a classification algorithm on the joint representation to generate labelled data from the joint representation, wherein the classification algorithm is trained with the input training data,
wherein the trained encoder comprises a different bias for each of the number of different modalities of input training data used to train the trained encoder.
15. The method of claim 14 , wherein only one modality of input data is provided.
16. The method of claim 14 , wherein performing classification comprises generating an indication of one or more regions of interest in the unlabelled data.
17. The method of claim 16 , wherein the one or more regions of interest are indicative of a cancerous growth.
18. The method of claim 14 , wherein encoding the unlabelled input data comprises encoding using a plurality of trained encoders.
19. The method of claim 14 , wherein the encoder of the multimodal neural encoder is a convolutional neural network, including any of: VGG neural network, AlexNet neural network, and/or recurrent neural network (RNN), optionally including bidirectional long short-term memory (LSTM) with 512 hidden units.