RU2822867C1 - Medical decision support hardware and software system - Google Patents

Abstract

FIELD: medicine.

SUBSTANCE: invention relates to a complex for diagnosing pulmonary pathologies. Complex comprises interconnected memory, processing device, communication unit for interaction with external databases and medical systems, service module containing: configuration and utilities module, preprocessing module, logic output module, interface module, detection module, wherein the service module, when receiving a request for processing a CT image for diagnosing pulmonary pathologies, initiates operation of the configuration module and utilities and sends it a CT image of the patient; configuration and utilities module connected to preprocessing module and interface module with specialized core, prepares input and output data of CT for transmission to preprocessing module and prepares file for configuration of neural network models for sending to interface module; preprocessing module limits the value of the data array of the CT image to the range of values [−1024;800], normalizes it in the range of [0;1], generates binary masks for three pathologies and sends the obtained data to the inference module; logical output module segments the CT image by means of a neural network located in the interface module, with a specialized core and sends the obtained data together with three binary masks to the detection module; detection module combines the initial CT image obtained from the inference module with three masks, by coloring the image with selection of pathologies and forming an array of data for displaying to a medical worker or for sending to an external medical system.

EFFECT: higher accuracy of processing data of CT images by means of simultaneous detection of three pathologies of lungs.

1 cl, 1 dwg

Description

FIELD OF THE INVENTION

The present invention relates to a hardware and software complex for supporting medical decision making using an artificial intelligence model based on a specialized Russian-made core, intended for analyzing computer topography images (CT images).

BACKGROUND OF THE ART

Coronavirus disease 2019 (COVID-19), an infectious disease that has resulted in a global pandemic and the most pressing threat facing humanity.

In this regard, the problem of early and accurate diagnosis of the presence of COVID disease becomes urgent. The most rapid method of identifying this disease is computed tomography (CT) of the lungs. Radiation methods, although not the main ones in diagnosing coronavirus infection, have become the most informative for identifying the presence and severity of changes in the lungs.

Consequently, today there is an urgent issue in the development of technical means for analyzing CT images in order to identify pathologies after COVID-19 lung damage, which have a significant impact on the treatment of complications or on making an alternative diagnosis.

In the prior art, there is a known source of information CN 112633404 A, disclosing a device for classifying CT images based on DenseNet for patients with COVID-19, which includes a data storage unit, a data preprocessing unit, a feature extraction unit, a unit for obtaining classification results, a receiving unit activation cards, display unit and control unit. This device processes CT images and displays the classification probability prediction value and ground glass pathology area to help the doctor make a diagnosis.

The disadvantage of this device is the lack of a more accurate diagnosis, in particular, the inability to determine such pathologies caused by coronavirus infection as: consolidation, pleural effusion, as well as the lack of operation of the device based on a Russian-made core. In addition, in this solution, the final diagnosis is made by the doctor, while the classification device predicts only the probability of the presence of pathology.

In the prior art, the source of information US 2022/0059221 A1 is known, disclosing a device operating on the basis of neural networks that processes CT images in order to determine the need for oxygen therapy. The disadvantage of this device is the lack of detection of pathologies caused by coronavirus infection, such as consolidation, pleural effusion, as well as the lack of operation of the device based on a Russian-made core.

Thus, the objective of the present invention is to develop a technical tool operating on the basis of a Russian-made core, which promptly detects pathologies caused by various diseases, in particular COVID, thereby reducing the time of patient admission, which provides highly qualified medical care and modern diagnostics to remote reception rooms patients in which medical personnel do not have the necessary expert qualifications to analyze CT results.

The technical result achieved by the claimed solution is the elimination of errors associated with the human factor that may occur in making a diagnosis.

An additional technical result is to increase the accuracy of CT image processing for timely detection of pathologies.

Another additional technical result should be considered the creation of a hardware and software complex operating on the basis of a Russian-made core, which promptly detects pathologies caused by various diseases.

DISCLOSURE OF INVENTION

The goal, the technical result required and obtained when using the invention is achieved by the fact that the hardware-software complex for diagnosing lung pathologies contains interconnected memory, a processing device, a communication unit for interacting with external databases and medical systems, a service module containing: a configuration module and utilities, preprocessing module, logical inference module, interface module, detection module, and:

a service module, which, when receiving a request to process a CT image for diagnosing lung pathologies, initiates the operation of the configuration module and utilities and sends it a CT image of the patient;

the configuration and utilities module, connected to the preprocessing module and the interface module with a specialized core, prepares the input and output CT data for transmission to the preprocessing module and prepares a file for configuring neural network models for sending to the interface module;

The preprocessing module limits the value of the data array to the range of values [-1024; 800], scales it in the range [0;1], generates binary masks for three pathologies and sends the received data to the logical inference module;

the logical inference module segments the CT image using a neural network located in the interface module with a specialized core and sends the received data in place with three binary masks to the detection module;

The detection module combines the original CT image obtained from the inference module with three masks, colorizing the image to highlight areas of pathology and forming a data array for display to a medical professional or for sending to an external medical system.

BRIEF DESCRIPTION OF THE DRAWINGS

The essence of the invention is illustrated by drawings.

Figure 1 shows a diagram of the architecture of the hardware and software complex.

IMPLEMENTATION OF THE INVENTION

The present invention provides a hardware and software complex for supporting medical decision making using an artificial intelligence model based on a specialized Russian-made core.

This complex ensures the determination of the presence or absence of signs of pathologies caused by COVID in the patient, such as: ground glass; consolidation zone; pleural effusion.

The hardware and software complex is deployed at the workplace in the operating organization and is connected to the Internet to exchange information with external medical systems and third-party databases using REST-API technology, and is also connected to the power supply network for operation around the clock.

The hardware and software complex contains interconnected memory, a processing device, a communication unit for interaction with external databases and medical systems, a service module containing: a configuration and utilities module, a preprocessing module, a logical inference module, an interface module, and a detection module.

In one embodiment, each of the configuration module and utilities, preprocessing module, logical inference module, interface module, classification and detection module is a computer program, each of which is located in a separate memory cell and is implemented by the processing device when it accesses the corresponding cell.

In another embodiment, each configuration and utility module, preprocessing module, inference module, interface module, classification and detection module is an arithmetic logic device.

The hardware and software complex operates on a Russian-made core

When a patient enters a medical institution (for example, a hospital), clinical data is collected, such as gender, age, blood test, CT images, etc., which are stored in a database.

The service module of the hardware and software complex, connected to the database via the Internet, initiates a request to the database to obtain a CT image of the patient.

In another embodiment of the invention, the hardware and software complex receives CT images from a computed tomography device to which it is connected via a network.

In another embodiment of the invention, the hardware and software complex receives CT images from external medical systems via a designated port using the REST-API service.

The service module is designed to organize the reception of requests and transmission of responses to external information systems and additionally performs the following functions:

- starts the process of listening to the input data port;

- processes requests;

- initializes and logs the operation of its modules in order to subsequently identify errors in their operation.

When receiving a request to process a CT image, the service module initiates the operation of the configuration module and the utility sends it a CT image of the patient.

The configuration and utilities module is designed to prepare input and output CT data for transmission, as well as to configure neural network models.

The CT data prepared by the configuration and utilities module in the form of a DICOM image is transferred to the CT results preprocessing module, in parallel with this, a configuration file in JSON format containing: name - the name of the CT model (neural network); width - width of the input image; height - height of the input image; num_threads - number of processor threads used; backend - library for executing the CT model (neural network), can take tensorflow or cuda values, is transferred to the interface module with a specialized kernel, which is part of the logical inference module.

The preprocessing module implements a two-stage processing algorithm.

The first of them is a preprocessing stage to obtain intervals of values of the attenuation coefficient, expressed in Hounsfield units for each class of pathologies (ground glass, consolidation zone, pleural effusion) [μ 1.5*σ; μ+1.5*σ], where:

μ is the mathematical expectation of the Hounsfield scale values (1) for voxels of each pathology class across the entire training data set, determined by the formula

where x _i is the value of the Hounsfield scale for the i-th voxel; N is the number of voxels belonging to a given pathology class in the entire data set; j=[1,N];

σ is the standard deviation of the Hounsfield scale values (2) for voxels of each pathology class across the entire training data set:

- D dispersion of Hounsfield scale values (3) for voxels of each pathology class across the entire training data set:

where x _i is the Hounsfield scale value for the i-oro voxel; N is the number of voxels belonging to a given pathology class in the entire data set; j=[1,N];

The second stage involves the formation of a binary mask for each of the three pathologies - an array equal in size to the CT slice (512x512x1), normalization of the CT data array and concatenation.

1. Formation of a binary mask for each of the three pathologies - an array equal in size to the CT slice (512x512x1).

To do this, the brightness value of the pixels on each slice is determined and compared with the intervals determined at the first stage.

If the brightness value of the corresponding pixel in the CT slice falls within the range of the interval previously formed for this pathology, then the element of the new array is assigned the value 1, if the brightness value of the corresponding pixel in the CT slice does not fall within the range of the interval previously formed for this pathology, then the element of the new array is assigned value 0.

The result is three binary masks: a binary mask consisting of 0 and 1, for pathology ground glass; a binary mask consisting of 0 and 1, for the pathology zone consolidation; a binary mask consisting of 0 and 1 for the pathology pleural effusion.

2. Normalization of the CT data array (voxels in the case of 3D, pixels in the case of 2D).

Raw CT image data values are expressed in Hounsfield units (HU), which means they are linearly normalized based on the attenuation coefficients of x-rays in water and air. However, the HU unit is not optimal for lung CT because mean lung CT values vary across data sets. Based on background information about the physical nature of the Hounsfield scale, intervals are assigned to the region in which tissues of interest may be located [-1024; 800].

Based on this interval, the data array containing CT data for each pathology is normalized:

if the value of the data array is less than the lower bound value, then it is assigned the value -1024; if the value of the CT data array is greater than the upper limit, then it is assigned 800.

Next, 1024 is added to each element and divided by 1824 to bring the value of all elements into the range [0;1].

3. Concatenation.

The resulting CT data array in the range [0;1] with a size of 512x512x1 and the three resulting binary masks with a size of 512x512x1 each are concatenated along the last (third) axis:

binary pathology mask #1 size [512x512x1] + binary pathology mask #2 size [512x512x1] + binary pathology mask #3 size [512x512x1] + CT image size [512x512x1] = array size [512x512x4]

The obtained values of all elements of CT images in the range [0;1] and binary masks for three pathologies (ground glass, consolidation zone, pleural effusion) are transferred to the inference module, in which the CT image is segmented by training a neural network, information about the configuration of which is located in the interface module with a specialized core.

The interface module with a specialized core organizes interaction with the Russian-made core and performs the following functions:

initializes the CT model with the following data: the number of processor threads used; CT model name; model execution library used;

supplies the input tensor to the input of the CT model;

returns an array of masks for each disease sign to the detection module.

The interface module is implemented as an engine for executing a neural network model on a Russian-made kernel and uses the functionality of the library supplied with the kernel.

The detection module combines a normalized CT image obtained from the inference module with three binary masks, thus colorizing the image (CT image) with highlighting areas of pathology, forming a data array for display to a medical professional or for sending to an external medical system. The data array (response) contains the final image in DICOM format. In case of a request processing error, the response contains: error class code; description of the error class for the corresponding error code; extended description of the error, which is logged by the service module.

As follows from the description of the claimed invention, the proposed hardware and software complex ensures the achievement of the stated technical result, namely, it ensures an increase in the accuracy of CT image data processing through the simultaneous determination of three pathologies caused by coronavirus infection: ground glass, consolidation, pleural effusion, thereby ensuring the exclusion errors associated with human factors that may occur in making a diagnosis. In addition, the special effectiveness of the hardware and software complex is its operation on the basis of a Russian-made core, which promptly detects pathologies caused by various diseases.

The analysis shows that all the general and specific features of the group of inventions are essential, since each of them is necessary for the industrial implementation of the claimed group of inventions, and together the features make it possible to achieve the declared technical result.

Thus, in the sections “State of the art” and “Disclosure of the invention” it was shown that all general and specific features of the proposed solution are essential and together satisfy the patentability criterion of “novelty”. In the section “Implementation of the invention” it was shown that the claimed group of inventions is technically feasible, allows solving the stated inventive problems and confidently achieving the required technical result when using it (the invention), which indicates the industrial applicability of the proposed solutions. Based on the above, we believe that the declared hardware and software complex for supporting medical decision-making satisfies all the requirements for patentability for inventions.

Claims (6)

A hardware and software complex for diagnosing lung pathologies, containing interconnected memory, a processing device, a communication unit for interaction with external databases and medical systems, a service module containing: a configuration and utilities module, a preprocessing module, a logical inference module, an interface module, a module detection, and: a service module, which, when receiving a request to process a CT image for diagnosing lung pathologies, initiates the operation of the configuration module and utilities and sends it a CT image of the patient; the configuration and utilities module, connected to the preprocessing module and the interface module with a specialized core, prepares the input and output CT data for transmission to the preprocessing module and prepares a file for configuring neural network models for sending to the interface module; The preprocessing module limits the value of the CT image data array to the range of values [-1024; 800], normalizes it in the range [0;1], generates binary masks for three pathologies and sends the received data to the logical inference module; the inference module segments the CT image using a neural network located in the interface module with a specialized core and sends the received data along with three binary masks to the detection module; The detection module combines the original CT image obtained from the inference module with three masks, colorizing the image to highlight areas of pathology and forming a data array for display to a medical professional or for sending to an external medical system.