FULLY CLOSED-CIRCUIT MEDICAL IMAGE DATA PROCESSING AND REPORT DRAFT GENERATION SYSTEM AND METHOD
Patent Information
- Application Number
- TR202612659
- Authority / Receiving Office
- TR · TR
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-09-21
Smart Images
Figure 00000011_0000
Abstract
Description
1 TARIFF FULLY CLOSED-CIRCUIT MEDICAL IMAGE DATA PROCESSING AND REPORT DRAFT PRODUCTION SYSTEM AND METHOD TECHNICAL AREA 5 The invention relates to medical image processing and computer-assisted clinical reporting systems. It relates to the technical field. In particular, the invention concerns three-dimensional images obtained from medical images. Spatial anomaly masking with voxel / pixel data, physically isolated. and 10 in the form of a draft report by a quantized local large language model It relates to a completely closed-loop system and method for processing. STATE OF THE ART In the current state of the technique, anatomical structures can be visualized on radiological images. or image processing and segmentation systems that perform anomaly detection 15 These systems can process an image in two or three dimensions. It can create labeling, segmentation maps, or anomaly masks. Together, the segmentation output is mostly in the form of images or numerical data. its retention requires further interpretation in the clinical reporting process and It requires conversion to text. 20 The current state of the art also includes natural language processing and large language modeling. Medical reporting systems that utilize external networks are available. Some of these systems operate via external networks. some of them utilize cloud-based computing resources accessed through it, while others... It generates report text by converting the image processing results into a language model. Transferring health data outside the institution, dependence on network connectivity and image 25 The difference in data structure between the processing output and the language model input affects these systems. Technical limitations in its implementation in closed infrastructures within hospitals. It constitutes. In the patent document with publication number CN 119092032 A, pre-processed Performing lesion recognition and segmentation on captured medical images, 30 Focal location from the segmentation map of the marked anomalous region, a pre-textual command containing information on morphological features, lesion type, and clinical significance obtaining and using a pre-trained large language model of the command in question It is explained that this process is carried out by [the relevant authority] and a medical examination report is generated. 2 In the patent document with publication number EP 4685808 A1, radiology images are described. By processing with artificial intelligence units, spatial and information regarding anatomical structures and pathologies can be obtained. generating explanatory data, preparing the initial radiology report, and through a user interface that includes image and text panels, the user It is stated that it will be submitted for review. 5 The aforementioned documents provide textual reporting of segmentation outputs. It includes solutions regarding its use. However, medical imaging Data transfer between the source and the local server is done using a hardware data diode. restricted by direction, the operation of the local server's network interfaces that go to the external internet The system is locked at the kernel level, image processing and language model 10 microservices are isolated from each other and the volume in the three-dimensional mask data, location and distribution values through local return-increasing production process a demand for an integrated data processing architecture that is translated into a quantized local language model Technical needs persist. THE PROBLEM THAT THE INVENTION AIMED TO SOLVE The technical problem that the invention aims to solve is medical image processing. resulting three-dimensional voxel / pixel data and spatial anomalies a local language whose masks cannot be processed directly due to data format differences Converting the model into a specific and repeatable input structure, this conversion and 20 Hardware and safety issues arise when patient data is leaked to the external network during the production of the draft report. a language that is blocked at the system level and requires high processing resources The model is run on local equipment within the hospital. Another technical problem is that the image processing unit and the language model unit are one and the same. The 25 components are run in isolation from each other without being tied to a single software structure. data transfer between units involves volume, position, and in three-dimensional mask data. through a standard intermediate data structure that preserves diffusion values It is the realization of. BRIEF DESCRIPTION OF THE INVENTION 30 The system described in the invention consists of a medical imaging source and local imaging from this source. A hardware data diode that creates one-way data transfer to the server side, local server. Image processing / segmentation microservice running on it, converter bridge, 3 The quantized local language model includes a microservice and a physician approval interface. Local The server's network interfaces that connect to the external internet are configured at the operating system kernel level. It is locked. The three-dimensional image produced by the image processing / segmentation microservice Spatial anomaly mask with voxel / pixel data, volume by converter bridge, 5 It is converted into structured data containing location and distribution values. Structured data, incremental generation of payback performed on the local server. It is converted into a deterministic command vector form through this process and quantized into a local language. The model is converted into a clinical report draft by the microservice. Image processing and language model microservices are isolated from each other and 10 Its containerized structure means that these units can operate independently of each other. It allows for modification. Hardware data diode and locked external network. The interfaces ensure that health data remains within the boundaries of the local system, while quantizing it. The language model reduces the memory and processing load on the local hardware. The generated The output is a draft report submitted for physician review in place of a diagnosis or treatment decision. 15 LIST OF FIGURES Figure 1 shows the draft of the fully closed-loop medical image data processing and reporting system, which is the subject of the invention. This is a schematic view illustrating the data flow architecture of the production system. The corresponding reference numbers used in the figures are: 1. Medical imaging source 2. Hardware data diode 3. Image processing / segmentation microservice 4. Converter bridge 25 5. Quantized local language model microservice 6. Physician approval interface DETAILED DESCRIPTION OF THE INVENTION The invention enables the transfer of medical image data to an off-site processing source. A fully closed-loop system for processing without transmission, and on this system... This is the data processing method used. Figure 1 shows the system's use in medical imaging. data flow architecture from source (1) to physician approval interface (6) It is shown. 4 The system includes at least one medical imaging source (1). Medical The imaging source (1) is a magnetic resonance device in an application. Medical Images taken from the imaging source (1) are two-dimensional pixel arrays or three-dimensional pixel arrays. The data processed within the scope of the invention may be in the form of three-dimensional voxel volumes. It is medical image data obtained first by the imaging device. The system, 5 It does not perform diagnostic or therapeutic procedures on human or animal bodies. Hardware data between medical imaging source (1) and local server The diode (2) is installed. The hardware data diode (2) is used for medical imaging. allowing data transfer from the source (1) to the local server and from the local server Reverse data flow to the medical imaging source or the network on the source side 10 It is a one-way network layer that prevents data transmission at the hardware level. Thus, data The direction of transfer is a physical connection, independent of a software access rule. It is limited by its regulations. The local server processes medical image data without using an external network connection. It is edge computing hardware. The network interfaces of the local server that connect to the external internet are operating system 15 The system is locked at the kernel level. The network is connected via a hardware data diode (2). By locking the interfaces together, data entry into the system and within the system are facilitated. Data processing is separated into different types: local server and hospital server. It could be an industrial edge computing hardware device that includes an external graphics processing unit. It can also be implemented in this way. 20 Image processing / segmentation microservice on local server (3), There is a translator bridge (4) and a quantized local language model microservice (5). Quantized local language model with image processing / segmentation microservice (3) Microservices (5) are isolated and containerized services. Within this structure Each microservice has its input and output data defined, and a microservice can have 25 modifying, changing the model structure of, other microservices It is not required. Image processing / segmentation microservice (3), from hardware data diode (2) deep learning-based image processing of acquired medical image data and It operates using a segmentation model. As a result of the process, 30 is obtained corresponding to the image volume. masking at least one spatial anomaly with the incoming three-dimensional voxel / pixel data. is created. The spatial anomaly mask is created within the image coordinate system. It represents the region marked as an anomaly. The converter bridge (4) is the digital image processing / segmentation microservice (3) between the output and the textual data input of the quantized local language model microservice (5) It is an intermediate data processing unit that eliminates the difference in data format. The converter bridge (4), spatial 5 The labeled voxel / pixel values in the anomaly mask and their corresponding images. By reading the coordinates, the volume, location, and distribution values of the anomaly can be determined. It determines the volume, location, and spread values, which have a fixed area sequence. It is embedded within structured data. Structured data generated by the translator bridge (4), return 10 It is used as query data in incremental production processes. In the payback process. The data used is stored within the boundaries of the local server and is not accessed by an external application. No data is being sent to the programming interface. The contextual data retrieved includes volume, location, and spread values. Structured data is converted to a deterministic instruction vector while preserving the specified field order. It is being converted. The command is for the same mask data and the same local contextual data. The field order of the vector and the data representation remain the same. The translator bridge (4) does not produce clinical diagnosis or report text; segmentation its output is a data structure that can be processed by the quantized local language model microservice (5) It transforms. 20 The input and output data structure of the converter bridge (4) must be specified, image Quantized local language with the model used in the processing / segmentation microservice (3). the independent selection of the model used in the microservice (5) Therefore, the converter bridge (4) allows for a specific image processing model. or creates an interface that is not tied to a specific language model. 25 Quantized local language model microservice (5), from the translator bridge (4) It processes the deterministic instruction vector on the local server. Large language the parameters of the model have been subjected to quantization or parameter compression This is how the model's memory requirements and computational load are reduced. hospital server or edge computing hardware with external graphics processing unit 30 It becomes possible to work on it. 6 Quantized local language model microservice (5), in deterministic instruction vector Structured findings in the form of a draft clinical report as textual output. It transforms. Quantized local language model microservice (5) created by The draft clinical report is transferred to the physician approval interface (6). Physician approval interface (6) allows viewing the draft report and making user corrections on the draft report. It receives the approval input. The draft report is the final clinical report before physician approval. or is not an autonomous diagnostic outcome. The data bus in the system is respectively the medical imaging source (1), hardware data diode (2), image processing / segmentation microservice (3), converter bridge (4), The quantized local language model is in the form of a microservice (5) and a physician approval interface (6). 10 Within this sequence of the data path, raw medical image data is converted to a language model. It is not given directly; first to a three-dimensional spatial anomaly mask, then... Structured data and deterministic data including volume, location, and spread values. It is converted into a command vector. The invention method primarily uses 15 taken from a medical imaging source (1) Medical image data is transmitted locally in one direction via hardware data diode (2). is transferred to the server. Then the image processing / segmentation microservice (3), Spatial anomaly masking from three-dimensional voxel / pixel data from medical image data. It constitutes. The converter bridge (4) converts the volume, position and spread values in the mask data to 20 converting to structured data with a fixed field order and local recovery It generates a deterministic command vector through an incremental production process. Quantized local Language model microservice (5), the command vector in question to the draft clinical report It is converted and the draft report is transferred to the physician approval interface (6). In one sample application of the invention, a three-dimensional magnetic field of view of a brain region was used. Resonance images are obtained from a medical imaging source (1). Image data is transmitted to the local server only via the hardware data diode (2). It is being transmitted in that direction. The local server's network interfaces that connect to the external internet are locked. Therefore, image data is sent to a processing source outside the system. Not being sent. 30 Images transferred to the local server undergo image processing / segmentation. a three-dimensional image coordinates processed by the microservice (3) and associated with the image coordinates 7 It is transformed into an anomaly mask. The marked area is indicated via this mask. numerical values are obtained that represent volume, position within the image, and spread. is being done. The converter bridge (4) consists of the numerical values in question from certain fields. converting it into structured data and performing the retrieval on the local server 5 It creates a deterministic instruction vector through an incremental production process. Thus, three The three-dimensional mask data does not require the raw image to be directly fed into the language model. without, the data structure that the quantized local language model microservice (5) can process It is being transformed. Quantized local language model microservice (5), 10 in deterministic instruction vector It processes volume, location, and distribution information to create a draft clinical report. The generated draft is displayed in the physician approval interface (6); by the physician The system is reviewed, necessary corrections are made, and user approval is granted. The text produced up to this stage constitutes an autonomous diagnosis or a final clinical report. It is not a report in the traditional sense, but a draft report submitted for physician evaluation. The 15 described above... The application example is given to illustrate how the invention works. It does not limit the scope of protection of the invention. 25
Claims
8 REQUESTS 1. A fully closed-loop medical image data processing and report drafting process. It is a system, and its feature is; - at least one medical imaging source providing medical image data (1), 5 - one-way from medical imaging source (1) to local server hardware data diode (2) that generates data traffic, - on a local server, isolated from each other and containerized in the form of microservices; from hardware data diode (2) From the acquired medical image data, three-dimensional voxel / pixel data and 10 Image processing / segmentation that removes spatial anomaly mask microservice (3) and deterministic command vector clinical report quantized local language model microservice (5) which transforms into a draft, - volume, position, and spread values in the spatial anomaly mask and organize as structured data with a fixed field order and 15 the structured data in question is data stored on the local server. the production process with increased return carried out on it translator bridge (4) which creates a deterministic command vector by applying, - physician viewing clinical report draft and receiving user input Confirmation interface (6) 20 includes and operates the network interfaces of the local server that connect to the external internet. the system is locked at the kernel level and the medical bus display source (1), hardware data diode (2), image processing / segmentation microservice (3), translator bridge (4), quantized local language model microservice (5), physician approval interface (6) ranked 25 It is characterized by its creation.
2. Fully closed-loop medical image data processing and report drafting according to Claim 1. It is a production system, and its feature is; magnetic (1) of medical imaging source. It is characterized by being a resonance device.
3. Fully closed-loop medical image data processing and report drafting according to Claim 1 30 It is a production system, and its feature is the spatial anomaly of the converter bridge (4). the voxel / pixel values marked in the mask and their image coordinates It is characterized by its ability to process and generate volume, location, and distribution values. 9 4. Fully closed-loop medical image data processing and reporting according to claim 1 or 3. It is a draft generation system, the feature of which is that the translator bridge (4) is in the same spatial Anomaly mask and data representation with field order for the same local contextual data. It is characterized by its ability to generate an unchanging deterministic command vector.
5. Fully closed-loop medical image data processing and report drafting according to Claim 1. It is a production system, and its feature is that it uses image processing / segmentation microservices. (3) and the defined input and output of the quantized local language model microservice (5). connected and independent of each other via the converter bridge (4) over data structures It is characterized by its interchangeability.
6. Fully closed-loop medical image data processing and report drafting according to Claim 10 It is a production system, and its feature is that it uses an external graphics processing unit (GPU) from the local server. including and parameter compression of the quantized local language model microservice (5). The external graphics processing unit in question is the implemented large language model. It is characterized by its focus on the subject.
7. Fully closed-loop medical image data processing and report drafting according to Claim 1 15 It is a production system, and its feature is; physician approval interface (6), clinical report by receiving user correction and user approval inputs regarding the draft It is characteristic.
8. It is a fully closed-loop medical image data processing method, characterized by: 20 medical image data from medical imaging source (1) One-way through the hardware data diode (2) to the external internet Network interfaces are locked locally at the operating system kernel level. transfer to the server, - Medical image data, isolated and containerized on a local server. through the image processing / segmentation microservice (3) which works as 25 Spatial anomaly masking with three-dimensional voxel / pixel data removal, - volume, position, and spread values in the spatial anomaly mask Structured data with fixed field order by converter bridge (4) to be organized as, 30 - structured data, data stored on the local server by implementing the production process with increased return on investment Generation of a deterministic command vector, - Deterministic instruction vector, image processing / segmentation Quantized 5 working in isolation and containerized form from the microservice (3) Draft clinical report by local language model microservice (5) transformation and - Transfer of the draft clinical report to the physician approval interface (6) It is characterized by including steps.
9. According to claim 8, it is a fully closed-loop medical image data processing method, and 10 Its feature is the spatial anomaly mask of volume, position, and dispersion values. from the marked voxel / pixel values and their image coordinates It is characterized by its creation.
10. According to Claim 8, it is a fully closed-loop medical image data processing method, feature; user 15 regarding clinical report draft in physician approval interface (6) It is characterized by the process of making corrections and obtaining user consent inputs. 25