Method and system for improving CBCT image definition

Abstract

The invention relates to the technical field of intelligent medical treatment, in particular to a method and system for improving CBCT image definition. The method comprises the following steps: constructing a CBCT image description generation model; on the basis of the real-time CBCT image and the CBCT image description generation model, obtaining expert scores and expert evaluation embedding of the real-time CBCT image; obtaining expert evaluation based on expert evaluation embedding and a pre-trained decoder; obtaining an SSIM value and a PSNR value of the real-time CBCT image; fusing the SSIM value, the PSNR value, the expert score and the expert evaluation according to a predetermined format to generate standard definition evaluation; and inputting the standard definition evaluation and the basic information of the subject into a constrained large language model, obtaining guide information of the subject, and broadcasting the guide information of the subject to the subject by voice. According to the method, the subject is effectively guided to adjust the state to reduce motion artifacts, so that the CBCT image definition is remarkably improved under the condition of low radiation dose.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, specifically to a method and system for improving the clarity of CBCT images. Background Technology

[0002] In minimally invasive liver surgery, CBCT (cone-beam computed tomography) plays an indispensable supporting role in ensuring high precision and safety. It guides and monitors the surgical process and its progress; surgeons rely on CBCT imaging information to precisely plan and execute the procedure when using minimally invasive equipment. However, CBCT itself has inherent limitations due to low radiation dose, imaging geometry, data acquisition methods, and reconstruction algorithms, resulting in imaging quality that is closely related to patient cooperation. Figure 1 This is a schematic diagram of a CBCT image. (Example) Figure 1 As shown, during surgery, patients are required to cooperate as much as possible to ensure highly clear CBCT images. This is because patient behavior, including breathing, subtle movements, and subconscious actions, can significantly impact the accuracy of surgical planning and execution. For example, involuntary breathing movements or minor limb tremors can cause artifacts in CBCT images, interfering with the surgeon's accurate assessment of the surgical site, thus affecting surgical outcomes and even endangering patient safety. Therefore, it is necessary to design a system that enables continuous image acquisition and incorporates systems to improve imaging quality.

[0003] Currently, existing technologies for improving CBCT imaging to assist in minimally invasive liver surgery can be broadly categorized into three types: (1) Breath-holding technique. This technique requires the patient to briefly hold their breath during the scan. This method is more suitable for patients who can cooperate with instructions, especially adults and some well-cooperative children. By holding their breath, the image blurring caused by respiratory motion can be reduced to some extent, thus improving image quality. (2) Respiratory gating technique. This technique uses respiratory sensors to monitor the patient's respiratory cycle and collects data only at specific stages of the respiratory cycle, such as the end of expiration. This method helps reduce motion artifacts caused by breathing, thereby improving the clarity of CBCT images and providing more accurate image information for surgery. (3) Image processing technology. Software algorithms are used to attempt to correct artifacts caused by motion. Although these methods cannot completely eliminate the problems caused by motion, they can improve image quality to some extent, allowing doctors to observe the surgical site more clearly.

[0004] However, the above-mentioned existing technologies all have objective disadvantages to varying degrees: (1) When implementing the breath-holding technique, an additional person needs to be arranged to give instructions to the patient. However, in actual surgery, if the patient is not given detailed explanation and psychological comfort before the operation, the patient may fail to cooperate with the breath-holding during the operation. Once the patient cannot hold their breath accurately according to the instructions, the quality of CBCT images will decrease, affecting the smooth progress of the operation. (2) The respiratory gating technique has high requirements for the patient's respiratory rhythm and is only suitable for patients with stable and regular respiratory rhythms. Moreover, its implementation process is relatively complicated, requiring additional respiratory training, gating threshold setting and real-time verification. These additional steps not only increase the workload of surgical preparation, but also increase the overall surgical process time, prolong the time the patient spends on the operating table, increase surgical risks and patient discomfort. (3) The current development of image processing technology is relatively lagging behind, mainly because multi-center multimodal datasets have not been established. Due to the lack of sufficiently rich and comprehensive data support, software algorithms cannot be effectively implemented and applied in a short period of time, and it is difficult to play a timely role in improving image quality in clinical surgery. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for improving the clarity of CBCT images in order to solve at least one of the aforementioned technical problems. This method involves image vision, image description generation, large language models, and machine learning. It analyzes the clarity of real-time CBCT images using image recognition methods and generates image descriptions using a decoder. Then, the image clarity description is input into a large language model to generate instructions specific to the patient, which are finally played back to the patient via voice, providing a more reliable and efficient CBCT imaging assistance solution for minimally invasive liver surgery.

[0006] The present invention achieves the above objectives through the following technical solutions: A method for improving the clarity of CBCT images includes: A CBCT image description generation model is constructed; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. Acquire real-time CBCT images, and obtain expert scores and expert evaluation embeddings of the real-time CBCT images based on the real-time CBCT images and the CBCT image description generation model; Expert evaluations are obtained based on the expert evaluation embedding and a pre-trained decoder; wherein, the expert evaluation is a textual language in a preset format used to describe CBCT images, and the expert evaluation embedding is a low-dimensional digital vector obtained by processing the expert evaluation. The real-time CBCT image is processed based on the SSIM algorithm and PSNR algorithm to obtain the SSIM value and PSNR value; the SSIM value, the PSNR value, the expert score, and the expert evaluation are fused according to a predetermined format to generate a standard sharpness evaluation. The standard clarity evaluation and the examinee's basic information are input into a constrained large language model to obtain examinee guidance information, and the examinee guidance information is then read aloud to the examinee.

[0007] Furthermore, a CBCT image description generation model is constructed, including: Obtain the CBCT image recognition dataset; The CBCT images in the CBCT image recognition dataset are processed to obtain the standardized image descriptions corresponding to each CBCT image; The standardized image description is processed to obtain expert scores and expert evaluation embeddings; The CBCT image is used as input, and the expert scores and expert evaluations corresponding to the CBCT image are embedded as output. A CNN-transformer model is trained to obtain the CBCT image description generation model.

[0008] Further, the CBCT images in the CBCT image recognition dataset are processed to obtain standardized image descriptions corresponding to each CBCT image, including: Each CBCT image is described by an expert in the field according to a fixed format; wherein the expert description includes: clarity, expert evaluation; the expert evaluation includes questions and suggestions; The expert description is converted into a standardized image description in a fixed format of “Clarity: x, Problem: [...], Recommendation: [...]” using the prompt large language model generation algorithm; where x is the clarity score, with a value of [0-6].

[0009] Further, the standardized image description is processed to obtain expert scores and expert evaluation embeddings, including: The resolution is then subjected to LLM standardization to obtain expert scores; The expert evaluation is processed by a tokenizer to obtain the expert evaluation embedding.

[0010] Furthermore, the training process of the decoder includes: The expert evaluations are segmented and encoded using a tokenizer to generate a numerical sequence. The numerical sequence is trained with an autoencoder to compress the number of its independent features and generate the expert evaluation embedding. The expert evaluation embedding is used as input, and the numerical sequence is used as output to generate the decoder.

[0011] Further, obtaining expert evaluations based on the expert evaluation embedding and the pre-trained decoder includes: The expert evaluation is embedded into the decoder to obtain the numerical sequence; The numerical sequence is de-tokenized to obtain the expert evaluation.

[0012] Furthermore, the basic information of the examinee includes: gender, age, and body type; The constrained large language model is constrained to "only provide emotional comfort and basic nursing guidance" and generates guidance information based on a pre-built response database, prohibiting the output of diagnosis, treatment or content beyond its scope; The response database includes a standard reassurance script library and a library of nursing instructions reviewed by doctors.

[0013] Furthermore, obtain guidance information from the examinee, including: Extract subject status keywords from the standard clarity assessment and combine them with the subject's basic information to retrieve the most relevant benchmark guidance information from the response database; The baseline guidance information is fed into the constrained large language model as context, limiting the constrained large language model to generate the guidance information only based on this context; If the semantic similarity between the guidance information and the context is lower than a set threshold, a refusal to answer will be triggered, and the patient will be prompted by voice to contact medical staff.

[0014] A system for improving the clarity of CBCT images includes: The model building unit is used to build a CBCT image description generation model; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. A real-time acquisition unit is used to acquire real-time CBCT images and, based on the real-time CBCT images and a CBCT image description generation model, obtain expert scores and expert evaluation embeddings for the real-time CBCT images. An evaluation decoding unit is used to obtain expert evaluations based on the expert evaluation embedding and a pre-trained decoder; wherein the expert evaluation is a text language in a preset format used to describe CBCT images, and the expert evaluation embedding is a low-dimensional digital vector obtained by processing the expert evaluation. The index fusion unit is used to process the real-time CBCT image based on the SSIM algorithm and PSNR algorithm to obtain the SSIM value and PSNR value; and to fuse the SSIM value, the PSNR value, the expert score, and the expert evaluation according to a predetermined format to generate a standard sharpness evaluation. The guidance broadcasting unit is used to input the standard clarity evaluation and the examinee's basic information into a constrained large language model, obtain the examinee's guidance information, and broadcast the examinee's guidance information to the examinee via voice.

[0015] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method for improving the clarity of CBCT images as described above.

[0016] The beneficial effects of this invention are as follows: This invention offers the advantage of real-time capability, enabling timely and precise instructions to patients. It guides patients during CBCT examinations to obtain clinically optimal image clarity with the lowest possible radiation dose, providing a reliable basis for surgery. During the procedure, it not only dynamically guides surgical operations based on image clarity but also provides timely and effective psychological guidance to patients, alleviating their anxiety and significantly enhancing the safety and success rate of the surgery.

[0017] According to this invention, the patient's behavior during CBCT examination can be dynamically adjusted, optimizing examination posture and cooperation based on real-time feedback, significantly improving the accuracy of surgical imaging and effectively controlling surgical precision. Especially in minimally invasive liver surgeries with limited time, it can rapidly improve operational accuracy, providing a strong guarantee for surgical success.

[0018] This invention innovatively combines Large Language Modeling (LLM) with automatic image annotation to standardize image sharpness, offering a highly efficient and convenient solution. LLM and an automatic encoder are used to standardize CBCT image sharpness, ensuring the accuracy and consistency of evaluation results. Simultaneously, continuous monitoring of surgical image sharpness allows for the keen identification and even, to some extent, prediction of image blurring caused by patient psychological risks, enabling proactive countermeasures.

[0019] This invention can effectively assist in the smooth conduct of surgery, reduce the workload of medical staff in judging image clarity and guiding patient communication, and help reduce labor costs. Compared with most existing technologies that fail to provide real-time feedback to patients, this invention fills this gap and brings a brand-new technical approach and practical solution to the field of CBCT image-assisted surgery. Attached Figure Description

[0020] Figure 1This is a schematic diagram of a CBCT image; where, Figure 1 (A) is a clear CBCT image; Figure 1 (B) Figure 1 (C) Figure 1 (D) shows three CBCT images with artifacts; Figure 2 This is a flowchart of a method for improving the clarity of CBCT images according to one embodiment of the present invention; Figure 3 This is a schematic diagram of a system structure for improving the clarity of CBCT images according to one embodiment of the present invention; Figure 4 This is a schematic diagram illustrating a system application for improving the clarity of CBCT images according to one embodiment of the present invention. Detailed Implementation

[0021] The invention will now be discussed with reference to exemplary embodiments. It should be understood that the described embodiments are merely intended to enable those skilled in the art to better understand and thus implement the invention, and are not intended to imply any limitation on the scope of the invention.

[0022] As used herein, the term "comprising" and its variations are to be interpreted as open-ended terms meaning "including but not limited to". The term "based on" is to be interpreted as "at least partially based on". The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment".

[0023] Example 1

[0024] Figure 2 This is a flowchart illustrating a method for improving CBCT image clarity according to one embodiment of the present invention. Figure 2 As shown, according to one embodiment of the present invention, a method for improving the clarity of CBCT images includes the following steps: Step S102: Construct a CBCT image description generation model; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. Step S104: Obtain real-time CBCT images, and obtain expert scores and expert evaluation embeddings of real-time CBCT images based on real-time CBCT images and CBCT image description generation models; Step S106: Obtain expert evaluation based on expert evaluation embedding and pre-trained decoder; wherein, expert evaluation is a pre-formatted text language used to describe CBCT images, and expert evaluation embedding is a low-dimensional digital vector obtained by processing expert evaluation. Step S108: Process the real-time CBCT image based on the SSIM algorithm and PSNR algorithm to obtain the SSIM value and PSNR value; fuse the SSIM value, PSNR value, expert score and expert evaluation according to a predetermined format to generate a standard sharpness evaluation. Step S110: Input the standard clarity assessment and the examinee's basic information into the constrained large language model, obtain the examinee's guidance information, and then verbally broadcast the examinee's guidance information to the examinee.

[0025] This embodiment proposes a method to improve the clarity of CBCT images. First, a CNN-Transformer model is trained using a hospital-owned liver CBCT dataset to construct a CBCT image description generation model, enabling the input CBCT image to directly output expert scores (0-6) and corresponding expert evaluation embedding vectors. Then, the model is invoked on real-time acquired CBCT images to simultaneously obtain expert scores and expert evaluation embeddings. A pre-trained autoencoder decoder then restores the embeddings to textual expert evaluations. In parallel, the SSIM and PSNR algorithms are used to calculate objective clarity indices for the same CBCT image frame, and the SSIM value, PSNR value, and expert score are combined. The expert evaluation is concatenated in a fixed order of "SSIM value; PSNR value; expert score; expert evaluation" to form a standard clarity evaluation sequence. When input into the constrained large language model, this sequence simultaneously carries three complementary clarity information (SSIM value, PSNR value, expert score), enabling the large language model to both perceive pixel-level distortion and understand clinical semantics. Thus, under the guidance of the RAG knowledge base, it generates safe and executable reassurance or action instructions. Finally, the instructions are fed back to the patient via voice broadcast, guiding them to adjust their breathing, reduce movements, or relieve tension. This achieves real-time closed-loop optimization of "image-evaluation-instruction-behavior-image," and the termination conditions are controlled by the doctor throughout the process, ensuring a safe and efficient surgical procedure.

[0026] This invention integrates expert scoring and evaluation, objective clarity indicators, and generates guidance information using a constrained large language model, achieving real-time closed-loop optimization of "image-evaluation-instruction-behavior-image," thereby improving CBCT image clarity and ensuring surgical safety and efficiency.

[0027] According to one embodiment of the present invention, step S102 includes: Step S1021: Obtain the CBCT image recognition dataset; Step S1022: Process the CBCT images in the CBCT image recognition dataset to obtain the standardized image descriptions corresponding to each CBCT image; Each CBCT image is described by a technical expert in the field according to a fixed format; the expert description includes: clarity, expert evaluation; the expert evaluation includes questions and suggestions; The prompt large language model generation algorithm is used to convert expert descriptions into a fixed format of "clarity: x, problem: [...], suggestion: [...]" as a standardized image description; where x is the clarity score, with a value of [0-6].

[0028] Step S1023: Process the standardized image description to obtain expert scores and expert evaluation embeddings; The resolution is standardized using LLM and expert scores are obtained. The expert evaluations are processed using a tokenizer to obtain the expert evaluation embeddings.

[0029] Step S1024: Using the CBCT image as input and the expert score and expert evaluation embedding corresponding to the CBCT image as output, train the CNN-transformer model to obtain the CBCT image description generation model.

[0030] In this embodiment, in step S102, a CBCT image recognition dataset covering both sharpness and various artifacts is first acquired. Then, imaging experts provide frame-by-frame original descriptions of three elements: "sharpness 0-6 points, problem-area, and suggestion." The prompt engineering-driven large language model rigorously normalizes any free text into a fixed format of "sharpness: x, problem: [...], suggestion: [...]," thus obtaining an unambiguous standardized image description. This description is then LLM-standardized to extract a numerical score as the expert's score. This score is data-driven (the original score of 0-6 minus the dataset mean and divided by the dataset standard deviation) to eliminate differences between batches. The scale shift of the annotation ensures that the subsequent model input follows an N(0,1) distribution, improving training stability. At the same time, the question-suggestion text is completed with fixed length using a tokenizer and then fed into an autoencoder to compress it into a low-dimensional expert evaluation embedding, achieving dual encoding of semantics and numerical values. Finally, using CBCT images as input and parallel standardized expert scores and expert evaluation embeddings as output, a CNN-Transformer model is trained to learn to directly map image grayscale-texture features to a clinically readable and numerically stable evaluation space, completing the construction of the CBCT image description generation model and providing a decodeable and fusionable unified representation for subsequent real-time closed-loop processing.

[0031] This invention constructs an image description generation model that can map CBCT image features to a clinically readable evaluation space by collecting datasets, normalizing expert descriptions, performing dual encoding processing, and training a CNN-Transformer model, providing a unified representation for subsequent closed-loop processing.

[0032] According to one embodiment of the present invention, in step S106, the training process of the decoder includes: The expert evaluations are segmented and encoded using a tokenizer to generate a numerical sequence. The numerical sequence is trained with an autoencoder to compress the number of its independent features and generate an expert evaluation embedding. The expert evaluation embedding is used as input, and the numerical sequence is used as output to generate a decoder.

[0033] Expert evaluations are obtained based on expert evaluation embeddings and pre-trained decoders, including: The expert evaluation is embedded into the input decoder to obtain a numerical sequence; The numerical sequence is de-tokenized to obtain expert evaluations.

[0034] In this embodiment, the decoder employs a "compress-then-decode" self-supervised strategy: During the training phase, the tokenizer converts fixed-format expert evaluations into a fixed-length token sequence, which is then compressed by the autoencoder into a low-dimensional expert evaluation embedding containing only the "question-region-suggestion" specificity. Subsequently, the decoder is trained using the embedding as input and the original tokens as reconstruction targets, enabling the embedding to accurately reproduce clinical text, reducing the embedding dimension to 1 / 8 of the sequence, and achieving a reconstruction error ≤ 1 token. During the inference phase, the expert evaluation embedding output by the CNN-Transformer is directly fed into the decoder to reconstruct the numerical token sequence. After being mapped by the de-tokenizer, the "question-region-suggestion" text consistent with the original expert semantics is immediately generated, completing the lossless reverse conversion from compressed vectors to readable evaluations. This provides a compact, efficient, and high-fidelity text source for standard clarity evaluation.

[0035] This invention enables the decoder to accurately reproduce clinically readable descriptions from low-dimensional expert evaluation embeddings through a "compression-then-decompression" self-supervised training process. This reduces the output burden while ensuring high fidelity and low latency, providing a reliable text source for standard clarity evaluation.

[0036] According to one embodiment of the present invention, in step S110, the basic information of the examinee includes: gender, age, and body shape; The constrained large language model is constrained to "only provide emotional comfort and basic nursing guidance" and generates guidance information based on a pre-built response database, prohibiting the output of diagnosis, treatment or content beyond its scope; The response database includes a standard reassurance script library and a library of physician-approved nursing instructions.

[0037] Preferably, obtaining subject guidance information includes: Keywords related to the examinee's status are extracted from the standard clarity assessment and combined with the examinee's basic information to retrieve the most relevant benchmark guidance information from the response database. The baseline guidance information is fed into the constrained large language model as context, which restricts the constrained large language model to only generate guidance information based on this context. If the semantic similarity between the guidance information and the context is lower than a set threshold, a refusal to answer will be triggered, and the patient will be prompted by voice to contact medical staff.

[0038] In this implementation, before generating guidance information for the examinee, three basic information items—gender, age, and body shape—are collected. These, along with state keywords such as "tension" and "motion artifacts" extracted from the standard clarity evaluation, form a retrieval vector. The system uses a constrained large language model that "only provides emotional reassurance and basic nursing guidance" as its execution core. It performs vector similarity retrieval in a response database composed of a standard reassurance script library and a doctor-reviewed nursing instruction library to lock in the most relevant baseline guidance information. This information is then fed into the model as the sole context, thus strictly limiting the generation space to reviewed content and preventing diagnostic or treatment suggestions. If the semantic similarity between the model output and the context is lower than a set threshold, a rejection mechanism is immediately triggered, and a fixed voice prompt is played to prompt the patient to contact medical staff. Otherwise, the safety guidance is converted into a voice broadcast, achieving compliant, real-time, and personalized intraoperative interaction.

[0039] This invention combines basic information of the examinee with status keywords to retrieve baseline guidance, and uses a constrained large language model to generate compliance guidance information, triggering a refusal mechanism to ensure safety, thereby achieving personalized, real-time and secure intraoperative interaction.

[0040] Example 2

[0041] According to one embodiment of the present invention, a method for improving CBCT clarity includes: Step S201: Obtain the CBCT image recognition dataset; The CBCT image recognition dataset includes clear CBCT images and CBCT images with various artifacts, with the proportions of different image types closely resembling real clinical acquisition conditions. The dataset contains both clear and unclear images; unclear images are caused by various patient movement patterns. These unclear images can be caused by: subtle patient movements (head-to-foot sliding, rotation / tilt), coughing, hiccups, swallowing, etc.; diaphragmatic-chest wall movements; and body tremors caused by tension and anxiety.

[0042] Step S202: Process the CBCT images in the CBCT image recognition dataset to obtain the standardized image descriptions corresponding to each CBCT image; For each CBCT image, an expert description is generated, and the image description is formatted and standardized via command-line engineering to obtain a standardized image description; specifically including: Step 1: Technical experts in the field provide an expert description for each CBCT image in a fixed format. The expert description includes: clarity (value [0-6]), problems, and suggestions. Step two: Standardize the expert descriptions to generate standardized image descriptions; Use prompt-engineering to standardize expert text into a formatted sentence like "Clarity: x, Problem: [...], Recommendation: [...]".

[0043] Step 3: Standardize the image descriptions as text labels and store the CBCT images and their corresponding text labels.

[0044] Clarity is rated on a scale of 0 to 6, with 0 indicating very poor and 6 indicating excellent.

[0045] The questions should be formatted as [Question 1 (area), Question 2 (area)] and may include 1-3 key questions. Each question must include a specific area (format: "Question Description (area)", such as "High Noise (Left Side of Liver Area)"), separated by commas; When describing regions, follow these guidelines: prioritize anatomical regions (left side of the liver / right side of the liver / gallbladder region), secondarily use image locations (upper center of the image / lower part of the image), and strictly avoid vague descriptions (such as "certain regions").

[0046] Recommendations should be brief clinical suggestions, such as "Re-collection is recommended" or "No treatment required".

[0047] For example: Example 1 – Clarity: 3, Problem: High noise (left side of liver area), blurred edges (upper center of image), Recommendation: CBCT fine-tuning recommended.

[0048] Example 2 – Sharpness: 6, Issues: None, Recommendation: No action required.

[0049] Example 3 – Sharpness: 2, Problems: Motion artifacts (right side of the liver area), low contrast (upper part of the image), Recommendation: Patient should reduce the range of motion.

[0050] Example 4 – Clarity: 4, Problem: Artifacts (gallbladder area), Recommendation: Patient should relax.

[0051] Command-line engineering can generate standardized image descriptions, which can then be used for subsequent automatic annotation model training for the sharpness module.

[0052] Step S203: Process the standardized image description to obtain expert scores and expert evaluation embeddings; The standardized image descriptions are encoded and quantified to obtain expert scores and expert evaluation embeddings; specifically including: Step 1: Obtain the sharpness in the standardized image description, perform LLM standardization on the sharpness, and obtain expert scores; Step 2: Obtain expert evaluations (including questions and suggestions) from standardized image descriptions, and encode the expert evaluations in segments using a tokenizer (Python library); When segmenting and coding expert evaluations, questions are filled into three categories; if there are fewer than three questions, they are left blank. Step 3: Train an autoencoder on the numerical sequence to compress the number of independent features and generate an expert evaluation embedding. Use a tokenizer to break down the standard sentence (clarity description) into a token→id sequence (expert evaluation embedding).

[0053] Step S204: Train and generate an autoencoder based on the relationship between expert evaluation and expert evaluation embedding; Train an autoencoder (compression → embedding) using the ID sequence, preserving both the encoder and decoder.

[0054] We construct an encoder and decoder pair, which, in conjunction with classic algorithms, can encode expert evaluations (including questions and suggestions) into expert evaluation embeddings through numerical encoding, and then decode them. The steps are as follows: First, we use a tokenizer (Python library) to segment and encode the expert evaluation descriptions, padding the questions to three; if there are fewer than three questions, leave them blank. Then, we train an autoencoder on the numerical sequence to compress the number of independent features. Finally, we retain the encoder and decoder.

[0055] Step S205: Process the CBCT image using the SSIM algorithm and the PSNR algorithm respectively to obtain the first data (SSIM value) and the second data (PSNR value). SSIM and PSNR are well-known image sharpness evaluation algorithms, which are not the focus of this invention and will not be elaborated further here. Step S206: Generate a sequence based on the first data (SSIM value), the second data (PSNR value), expert scores, and expert evaluations as a standard sharpness evaluation; The image description embedding module acquires the image and then transforms it into a sequence, i.e., a standard sharpness evaluation: ; Step S207: Using CBCT images as input and expert scores and expert evaluation embeddings as output, train the CNN-transformer model to obtain the image description generation model; Convolutional Neural Networks (CNNs) are an important tool in image recognition. By recognizing image-related features, CNNs can extract features of blurred patterns in images. Transformers, on the other hand, can correlate the features of blurred image patterns with standard sharpness evaluations.

[0056] Specifically, the CNN-transformer model comprises two modules. The first is a CNN feature extraction module containing several (5-10 groups) of basic feature extraction units [convolution-normalization-activation-pooling]. The second part is a standard transformer structure, taking image features as input and expert scores and expert evaluation embeddings as output. Based on the correspondence between CBCT images and expert scores and expert evaluation embeddings, the CNN-transformer model is trained using training set data and validated using a test set. When the parameter set in the model satisfies the convergence condition, the final CNN-transformer model is obtained.

[0057] Step S208: Obtain real-time CBCT images of the patient, and obtain expert scores and expert evaluation embeddings based on the real-time CBCT images and image description generation model.

[0058] Expert scores and expert evaluation embeddings can be obtained based on real-time CBCT images and CNN-transformer models; Step S209: Based on expert evaluation of the embedding, decoder, and detokenizer, obtain expert evaluation (including questions and suggestions).

[0059] Step S210: Process the real-time CBCT image using the SSIM algorithm and PSNR algorithm to obtain first data and second data; stitch together the first data, second data, expert score (obtained in step S208), and expert evaluation (obtained in step S209) to obtain the standard sharpness evaluation. In step S211, the standard clarity assessment and the patient's basic information are input into the large language model. The large language model outputs comforting answers and conveys guiding instructions to the patient.

[0060] By combining standardized image descriptions, the Retrieval-Augmented Generation (RAG) method constrains the large language model to only respond to patient reassurance and generate instructions by integrating information retrieval with text generation. This effectively limits the application scope of the large language model to specific tasks in healthcare, ensuring that its output is both professional and safe. The cue word engineering is set as follows: the model is "a virtual nurse assistant focused on providing emotional support and basic nursing guidance," and it is explicitly stated that "you can only answer questions related to patient emotional comfort, psychological counseling, and instructions for executing doctor-prescribed orders. For any medical diagnoses, treatment suggestions, or non-medical questions outside this scope, you must refuse to answer and guide the patient to contact professional healthcare personnel." This strong constraint acts as the first line of defense, limiting the model's "thinking" direction and output content from the source. The Retrieval-Augmented Generation method constructs a dedicated knowledge base containing a "standard reassurance script library" and a "doctor-reviewed nursing instruction library." When a patient appears tense in an image, the RAG system first retrieves the most relevant reassuring phrases or standard operating instructions from this controlled knowledge base. It then provides the retrieved, precise, and safe content as context to the large language model, guiding it to generate the final response. This ensures that the model's output is based on pre-verified, accurate information, rather than relying on unverified general knowledge that may exist within its internal parameters, thus significantly reducing the risk of generating incorrect or harmful information. Finally, the generated content is played back via voice to interact with the patient, giving them instructions and improving image quality.

[0061] For example, when a patient hears instructions such as "Please exhale gently and relax your shoulders," they subconsciously reduce their movements and adjust their breathing, thus changing their posture. Once the next CBCT image is acquired, the system continues monitoring until the doctor deems the image clear enough to begin surgery or proceed to the next step. The termination condition of this cycle is controlled by the doctor.

[0062] Example 3 Figure 3 This is a schematic diagram of a system structure for improving the clarity of CBCT images according to one embodiment of the present invention; Figure 4 This is a schematic diagram illustrating a system application for improving CBCT image clarity according to one embodiment of the present invention. Figure 3-4 As shown, according to one embodiment of the present invention, a system for improving the clarity of CBCT images includes: Model building unit 10 is used to build a CBCT image description generation model; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. The real-time acquisition unit 20 is used to acquire real-time CBCT images and obtain expert scores and expert evaluation embeddings of real-time CBCT images based on real-time CBCT images and CBCT image description generation models. The evaluation decoding unit 30 is used to obtain expert evaluation based on expert evaluation embedding and a pre-trained decoder; wherein, the expert evaluation is a text language in a preset format used to describe CBCT images, and the expert evaluation embedding is a low-dimensional digital vector obtained by processing the expert evaluation. The index fusion unit 40 is used to process real-time CBCT images based on the SSIM algorithm and PSNR algorithm to obtain SSIM value and PSNR value; and to fuse the SSIM value, PSNR value, expert score and expert evaluation according to a predetermined format to generate a standard sharpness evaluation. The guidance broadcasting unit 50 is used to input the standard clarity evaluation and the examinee's basic information into the constrained large language model, obtain the examinee's guidance information, and broadcast the examinee's guidance information to the examinee via voice.

[0063] In this embodiment, a system for improving the clarity of CBCT images is proposed. The CNN-Transformer model trained offline by the model building unit 10 receives any CBCT frame and outputs expert scores and corresponding low-dimensional expert evaluation embeddings. The real-time acquisition unit 20 sends the expert evaluation embeddings and expert scores to the evaluation decoding unit 30 in real time. After two stages of reconstruction by the autoencoder decoder and the detokenizer, the plaintext "problem-region-suggestion" is obtained. The index fusion unit 40 calculates the SSIM value and PSNR value of the same frame in parallel, according to "SSIM value; PSNR value; The expert scoring and evaluation process is pre-formatted to form a standard-definition evaluation stream. The guidance broadcasting unit 50 combines this stream with basic patient information such as gender, age, and body shape, inputting it into a constrained large language model that "only provides emotional support and basic nursing guidance." It then retrieves the most relevant baseline guidance from the doctor-approved response database, generates compliant voice commands after similarity threshold verification, and plays them in real-time to guide patients in adjusting their breathing or posture. This completes the closed-loop optimization of image-evaluation-command-behavior, requiring no additional human intervention throughout the process. Furthermore, all units are deployed modularly and can be seamlessly integrated with existing CBCT workstations.

[0064] This invention achieves closed-loop optimization of image evaluation, indicator fusion, and compliant voice guidance through modular units. The entire process requires no human intervention and can be seamlessly integrated with existing equipment, effectively improving image clarity and ensuring intraoperative safety and efficiency.

[0065] According to one embodiment of the present invention, an electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements any of the methods of the present invention for improving the clarity of CBCT images.

[0066] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and equipment described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0067] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0068] It should be understood that the sequence number of each step in the invention and embodiments of the present invention does not absolutely imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

Claims

1. A method for improving the clarity of CBCT images, characterized in that, include: A CBCT image description generation model is constructed; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. Acquire real-time CBCT images, and obtain expert scores and expert evaluation embeddings of the real-time CBCT images based on the real-time CBCT images and the CBCT image description generation model; Expert evaluations are obtained based on the expert evaluation embedding and a pre-trained decoder; wherein, the expert evaluation is a textual language in a preset format used to describe CBCT images, and the expert evaluation embedding is a low-dimensional digital vector obtained by processing the expert evaluation. The real-time CBCT image is processed based on the SSIM algorithm and PSNR algorithm to obtain the SSIM value and PSNR value; the SSIM value, the PSNR value, the expert score, and the expert evaluation are fused according to a predetermined format to generate a standard sharpness evaluation. The standard clarity evaluation and the examinee's basic information are input into a constrained large language model to obtain examinee guidance information, and the examinee guidance information is then read aloud to the examinee.

2. The method for improving CBCT image clarity according to claim 1, characterized in that, Constructing a CBCT image description generation model includes: Obtain the CBCT image recognition dataset; The CBCT images in the CBCT image recognition dataset are processed to obtain the standardized image descriptions corresponding to each CBCT image; The standardized image description is processed to obtain expert scores and expert evaluation embeddings; The CBCT image is used as input, and the expert scores and expert evaluations corresponding to the CBCT image are embedded as output. A CNN-transformer model is trained to obtain the CBCT image description generation model.

3. The method for improving CBCT image clarity according to claim 2, characterized in that, The CBCT images in the CBCT image recognition dataset are processed to obtain standardized image descriptions corresponding to each CBCT image, including: Each CBCT image is described by an expert in the field according to a fixed format; wherein the expert description includes: clarity, expert evaluation; the expert evaluation includes questions and suggestions; The expert description is converted into a standardized image description in a fixed format of "Clarity: x, Problem: [...], Recommendation: [...]" using the prompt large language model generation algorithm; where x is the clarity score, with a value of [0-6].

4. The method for improving CBCT image clarity according to claim 2, characterized in that, The standardized image description is processed to obtain expert scores and expert evaluation embeddings, including: The resolution is then subjected to LLM standardization to obtain expert scores; The expert evaluation is processed by a tokenizer to obtain the expert evaluation embedding.

5. The method for improving CBCT image clarity according to claim 1, characterized in that, The training process of the decoder includes: The expert evaluations are segmented and encoded using a tokenizer to generate a numerical sequence. The numerical sequence is trained with an autoencoder to compress the number of its independent features and generate the expert evaluation embedding. The expert evaluation embedding is used as input, and the numerical sequence is used as output to generate the decoder.

6. The method for improving CBCT image clarity according to claim 5, characterized in that, Expert evaluations are obtained based on the expert evaluation embedding and a pre-trained decoder, including: The expert evaluation is embedded into the decoder to obtain the numerical sequence; The numerical sequence is de-tokenized to obtain the expert evaluation.

7. The method for improving CBCT image clarity according to claim 1, characterized in that: The basic information of the examinee includes: gender, age, and body shape; The constrained large language model is constrained to "only provide emotional comfort and basic nursing guidance" and generates guidance information based on a pre-built response database, prohibiting the output of diagnosis, treatment or content beyond its scope; The response database includes a standard reassurance script library and a library of nursing instructions reviewed by doctors.

8. The method for improving CBCT image clarity according to claim 7, characterized in that, Obtain guidance information for the examinee, including: Extract subject status keywords from the standard clarity assessment and combine them with the subject's basic information to retrieve the most relevant benchmark guidance information from the response database; The baseline guidance information is fed into the constrained large language model as context, limiting the constrained large language model to generate the guidance information only based on this context; If the semantic similarity between the guidance information and the context is lower than a set threshold, a refusal to answer will be triggered, and the patient will be prompted by voice to contact medical staff.

9. A system for improving the clarity of CBCT images, characterized in that, include: The model building unit is used to build a CBCT image description generation model; the input of the CBCT image description generation model is a CBCT image, and the output is an expert score and expert evaluation embedding used to describe the CBCT image. A real-time acquisition unit is used to acquire real-time CBCT images and, based on the real-time CBCT images and a CBCT image description generation model, obtain expert scores and expert evaluation embeddings for the real-time CBCT images. An evaluation decoding unit is used to obtain expert evaluations based on the expert evaluation embedding and a pre-trained decoder; wherein the expert evaluation is a text language in a preset format used to describe CBCT images, and the expert evaluation embedding is a low-dimensional digital vector obtained by processing the expert evaluation. The index fusion unit is used to process the real-time CBCT image based on the SSIM algorithm and PSNR algorithm to obtain the SSIM value and PSNR value; and to fuse the SSIM value, the PSNR value, the expert score, and the expert evaluation according to a predetermined format to generate a standard sharpness evaluation. The guidance broadcasting unit is used to input the standard clarity evaluation and the examinee's basic information into a constrained large language model, obtain the examinee's guidance information, and broadcast the examinee's guidance information to the examinee via voice.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method for improving the clarity of CBCT images as described in any one of claims 1-8.

Images