System, method and device for predicting and generating surgical intervention images
A system that includes data collection engine devices, client devices and backend devices. The backend devices include trained models, business logic, and attributes of a plurality of patient events. A plurality of data collection engines and hospital information systems send input attributes of new patient events to the backend devices. The backend devices can generate output images predicting particular outcomes of new patient events based upon the input attributes utilizing the trained models.
1. A computer implemented method for generating a medical output image associated with a new event, the method comprising:
storing a plurality of past events, each of the plurality of past events including a plurality of past input attributes and a quantifiable outcome; and
training a neural network model (NNM) to generate a trained model, wherein the training of the NNM includes:
performing pre-processing on the plurality of past input attributes for each of the plurality of past events to generate a plurality of past input data sets;
dividing the plurality of past events into a first set of training data and a second set of validation data;
iteratively performing a machine learning algorithm (MLA) to update synaptic weights of the NNM based upon the training data; and
validating the NNM based upon the second set of validation data, receiving a plurality of input attributes of the new event, the input attributes associated with a plurality of segmented images;
performing pre-processing on the plurality of input attributes to generate an input data set; and
generating the medical output image from the trained model based upon the input data set, the medical output image being a reconstruction generated from the plurality of segmented images.
2. The method of claim 1 , wherein:
the NNM includes an input layer, output layer, and a plurality of hidden layers with a plurality of hidden neurons; and
each of the plurality of hidden neurons includes an activation function, the activation function is one of:
(1) the sigmoid function ƒ(x)=1/(1+e −x );
(2) the hyperbolic tangent function ƒ(x)=(e 2x −1)/(e 2x +1); and
(3) a linear function ƒ(x)=x,
wherein x is a summation of input neurons biased by the synoptic weights.
3. The method of claim 1 , wherein the NNM is one or more of a feed forward structure Neural Network; Convolutional neural network (CNN); ADALINE Neural Network, Adaptive Resonance Theory 1 (ART1), Bidirectional Associative Memory (BAM), Boltzmann Machine, Counterpropagation Neural Network (CPN), Elman Recurrent Neural Network, Hopfield Neural Network, Jordan Recurrent Neural Network, Neuroevolution of Augmenting Topologies (NEAT), and Radial Basis Function Network.
4. The method of claim 1 , wherein the performing of the MLA includes measuring a global error in each training iteration for the NNM by:
calculating a local error, the local error being a difference between the output value of the NNM and the quantifiable outcome;
calculating the global error by summing all of the local errors in accordance with one of:
(1) Mean Square Error (MSE) formula
∑
n
E
2
n
;
(2) Root Mean Square Error (RMS) formula
∑
n
E
2
n
;
and
(3) Sum of Square Errors (ESS) formula
∑
n
E
2
2
,
wherein n represents a total number of the past patient events and E represents the local error.
5. The method of claim 1 , wherein the plurality of segmented images are from an imaging study.
6. The method of claim 1 , wherein
the medical output image is a 3-D reconstruction generated from the plurality of segmented images.
7. The method of claim 6 , wherein the imaging study includes a computed tomography (CT) scan of a head of the patient using series of x-rays of the head taken from many different directions, and each of the segmented images is a Digital Imaging and Communications in Medicine (DICOM) image.
8. The method of claim 1 , wherein:
the receiving of the plurality of input attributes of the new event further includes receiving a Digital Imaging and Communications in Medicine (DICOM) imaging file including patient identification and an array of pixel intensity data associated with a computed tomography (CT) scan;
the performing of pre-processing of the plurality of input attributes to generate an input data set further comprises windowing values of the array of pixel intensity data outside of a predetermined range; and
the generating of the medical output image from the trained model based upon the input data set further includes generating the medical output image based upon a Hounsfield unit (HU) pixel intensity value associated with the windowed values of the array of pixel intensity data.
9. The method of claim 7 , wherein the medical output image is a blood clot image when the HU pixel intensity value is greater than or equal to 50 and less than or equal to 75.
10. The method of claim 7 , wherein the medical output image is a subdural hematoma image when the HU pixel intensity value is greater than or equal to 75 and less than or equal to 100 within 24 hours from date of imaging study.
11. The method of claim 7 , wherein the medical output image is a subdural hematoma image when the HU pixel intensity value is greater than or equal to 65 and less than or equal to 85 within 72 hours from date of imaging study.
12. The method of claim 7 , wherein the medical output image is a subdural hematoma image when the HU pixel intensity value is greater than or equal to 35 and less than or equal to 40 within 10 days from date of imaging study.