Hybrid convolutional wavelet networks for predicting treatment response via radiological images of bowel disease

A method of image analysis, comprising:

• forming an imaging dataset comprising imaging data corresponding to one or more pre-treatment radiological images of a patient having a bowel disease;
• performing a convolution on the imaging data to form a feature map;
• performing a rectification function on the feature map to form a rectification map;
• performing a pooling operation on the rectification map to form a plurality of CNN outputs;
• applying a wavelet network to the plurality of CNN outputs, wherein the wavelet network is configured to decompose the plurality of CNN outputs into a plurality of different versions of a mother wavelet respectively having different wavelet shift parameters and wavelet scale parameters, and to generate a plurality of CWN outputs from the plurality of different versions of the mother wavelet;
• and
• constructing a predictive signature for the patient based on the plurality of CWN outputs, the predictive signature being associated with disease response or risk.

An image analysis system, comprising:

• a memory configured to store an imaging dataset comprising imaging data corresponding to one or more pre-treatment radiological images of a patient having a bowel disease;
• one or more convolutional neural network (CNN) segments configured to perform one or more convolutions on the imaging data to generate a plurality of CNN outputs;
• a wavelet network disposed downstream of the one or more CNN segments, the wavelet network being configured to decompose the plurality of CNN outputs into a plurality of wavelons and to generate a plurality of CWN outputs by summing differently weighted versions of the plurality of wavelons, wherein the plurality of wavelons respectively comprise different versions of a mother wavelet that has been shifted and scaled according to different wavelet shift parameters and different wavelet scale parameters;
• and
• a predictive signature generator configured to generate a predictive signature relating to the bowel disease of the patient.