Breast cancer doctor digital clone construction method based on three-layer corpus engineering, electronic device and computer readable storage medium

Abstract

The application discloses a breast cancer doctor digital clone construction method based on a three-layer corpus engineering, obtains a target doctor multi-view face image, performs three-dimensional reconstruction based on 3D Gaussian splashing to obtain a drivable Gaussian sphere representation, integrates the drivable Gaussian sphere representation with a lip-shaped synchronous model to realize audio-driven face animation generation, adopts a double-adapter low-rank adaptive fine-tuning architecture to train a loop evidence adapter and a style adapter respectively and controls the contribution proportion through a weighting coefficient, constructs a whitelist library, a blacklist library and a gray area problem library, and defines a comprehensive confidence index which is weighted fusion of a model self-confidence and a loop evidence confidence, triggers a fuse when the comprehensive confidence index is lower than a preset threshold, and applies an uncertainty penalty coefficient when a gray area problem is triggered, unifies an image layer, a corpus layer and a boundary layer in a single framework, and realizes collaborative gain of image reality, language individualization and safety controllability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, computer vision, and natural language processing, and specifically to a method for constructing digital avatars of breast cancer doctors based on three-layer corpus engineering, electronic devices, and computer-readable storage media. Background Technology

[0002] Breast cancer is the most common malignant tumor among women worldwide, and patients face significant information and emotional support needs throughout the entire treatment process. Existing medical AI systems suffer from structural flaws in the construction of digital avatars for doctors, with fragmented image, corpus, and boundary layers. At the image layer, current medical AI systems rely on pure text or voice interaction, lacking a visual representation of the doctor. 3D Gaussian splashing digital human technology with facial animation generation capabilities is geared towards general scenarios such as virtual reality, games, and video conferencing, and is not integrated with medical content generation systems. Existing digital human lip-sync technology only addresses the single technical aspect of audio-visual alignment and is not integrated into a complete medical digital avatar system architecture. At the corpus layer, existing medical AI systems use general medical knowledge bases to generate responses, resulting in a consistent response style across all patients and failing to reflect the personalized language characteristics of specific doctors. At the boundary layer, large language models are prone to generating erroneous content, and existing security solutions rely on keyword filtering or external multi-model voting architectures. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method for constructing digital avatars for breast cancer doctors based on three-layer corpus engineering, in order to solve the problem that the image layer, corpus layer and boundary layer are isolated from each other, making it impossible to unify the doctor's personalized image reconstruction, language style transfer and security boundary control into a single framework.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a digital avatar of a breast cancer doctor based on a three-layer corpus engineering, wherein the image layer is constructed by acquiring multi-view facial images of the target doctor, performing three-dimensional reconstruction based on 3D Gaussian splashing to obtain a driveable Gaussian sphere representation, and integrating the Gaussian sphere representation with a lip-sync model to achieve audio-driven facial animation generation. Corpus layer construction: A dual-adaptor low-rank adaptive fine-tuning architecture is adopted. The evidence-based adapter and the style adapter are trained separately on the pre-trained large language model. During the inference stage, the contribution ratio of the evidence-based adapter and the style adapter is controlled by weighting coefficients. Boundary layer construction: Construct a whitelist library, a blacklist library, and a gray zone problem library. Define a comprehensive confidence index that is a weighted fusion of the model's own confidence and the confidence of evidence-based evidence. When the comprehensive confidence index is lower than a preset threshold, a circuit breaker is triggered. When a gray zone problem is triggered, an uncertainty penalty coefficient is applied to the comprehensive confidence index.

[0005] Further optimizations of this invention include extracting language style features from four dimensions: lexical preferences, syntactic complexity, empathic expression, and risk communication patterns.

[0006] Further optimizations of this invention include mapping the uncertainty penalty coefficient to three levels of gray zone risk levels: the uncertainty penalty coefficient for low risk level is 0.85 to 0.95, the uncertainty penalty coefficient for medium risk level is 0.70 to 0.85, and the uncertainty penalty coefficient for high risk level is 0.50 to 0.70.

[0007] Further optimizations of this invention include the following: in the comprehensive confidence index, the model's own confidence is calculated based on the normalized information entropy of the probability distribution of the output sequence, and the evidence-based confidence is calculated based on the semantic retrieval matching degree between key medical assertions in the generated content and the evidence-based knowledge base.

[0008] Further optimizations of this invention include obtaining the target doctor's modifications to the draft response, recording the text pairs before and after the modifications, triggering incremental fine-tuning of the style adapter when the cumulative number of valid modification pairs reaches a preset threshold, freezing the weights of the evidence-based adapter and the base weights of the pre-trained large language model during fine-tuning, and updating only the low-rank matrix of the style adapter.

[0009] Further optimizations of this invention include performing INT8 quantization on the Gaussian sphere representation to reduce storage volume, removing Gaussian spheres whose contribution to the peak signal-to-noise ratio of the rendering is lower than a preset contribution threshold through structured pruning, training a lightweight student network, and learning from the original 3D Gaussian splash rendering output in a knowledge distillation manner to obtain an image layer inference model that can run in real time on mobile devices.

[0010] Further optimizations of this invention include, after incremental fine-tuning, evaluating the style similarity before and after the style adapter update through A / B testing, and automatically replacing the online style adapter once the style similarity improvement reaches a preset standard.

[0011] Further optimizations of this invention include organizing a collaborative workflow according to four stages: AI drafting, evidence verification, doctor review, and digital human broadcasting. In the AI ​​drafting stage, a style adapter generates a draft response that matches the target doctor's style. In the evidence verification stage, the evidence-based knowledge base is automatically searched to trace and annotate key medical assertions in the draft. In the doctor review stage, the verified draft is submitted to the target doctor for review and modification. In the digital human broadcasting stage, the final approved response is output in the form of lip-sync video.

[0012] An electronic device includes a processor and a memory, the memory storing instructions that, when executed by the processor, implement the above-described method for constructing a digital avatar of a breast cancer doctor based on three-layer corpus engineering.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for constructing digital avatars of breast cancer doctors based on three-layer corpus engineering.

[0014] The beneficial effects of this invention are: it unifies the image layer, corpus layer, and boundary layer under a three-layer corpus engineering framework, with synergistic gains among the layers. The image layer provides a sense of presence for the output of the corpus layer, the language style of the corpus layer gives the image layer's broadcasting the doctor's own linguistic characteristics, and the boundary layer provides security for the output of both the image layer and the corpus layer; A dual-adaptor decoupled architecture of evidence-based adapter and style adapter is adopted to separate medical knowledge injection and language style transfer into independent low-rank adaptive fine-tuning paths. During the inference stage, the contribution ratio of evidence-based adapter and style adapter is controlled by weighting coefficients to achieve an adjustable balance between medical accuracy and style fidelity. A three-level classification system is constructed, consisting of a whitelist, a blacklist, and a gray zone issue database. A comprehensive confidence index is defined, which is a weighted fusion of the model's own confidence and the confidence of evidence-based evidence. An uncertainty penalty coefficient is applied when a gray zone issue is triggered, and a circuit breaker is triggered when the comprehensive confidence index is lower than a preset threshold, thereby achieving fine-grained security control embedded in the generation process. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0016] Figure 2 This is a schematic diagram of the hierarchical architecture of the three-layer corpus engineering of the present invention.

[0017] Figure 3 This is a schematic diagram of the low-rank adaptive dual adapter decoupling architecture of the present invention.

[0018] Figure 4 This is a schematic diagram of the four-stage closed-loop process of human-machine collaboration according to the present invention.

[0019] Figure 5 This is a schematic diagram of the confidence-based circuit breaker decision-making process of the present invention.

[0020] Figure 6 This is a first enlarged view of the hierarchical architecture diagram of the three-layer corpus engineering of the present invention.

[0021] Figure 7 This is a second enlarged view of the hierarchical architecture diagram of the three-layer corpus engineering of the present invention.

[0022] Figure 8 This is a third enlarged view of the hierarchical architecture diagram of the three-layer corpus engineering of the present invention.

[0023] Figure 9This is a first partial enlarged view of the schematic diagram of the low-rank adaptive dual adapter decoupling architecture of the present invention.

[0024] Figure 10 This is a second enlarged view of the schematic diagram of the low-rank adaptive dual adapter decoupling architecture of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the embodiments shown in the accompanying drawings.

[0026] A method for constructing digital avatars for breast cancer doctors based on three-layer corpus engineering includes the following steps: S1 Image Layer Construction: Acquire multi-view facial images of the target doctor, perform three-dimensional reconstruction based on 3D Gaussian splash (3DGS) to obtain a driveable Gaussian sphere representation of the target doctor, and then integrate the Gaussian sphere representation with a pre-trained lip-sync model to realize audio-driven facial animation generation of the doctor.

[0027] S2 Corpus Construction: A training dataset containing medical knowledge corpus and doctors' personal corpus is constructed; a dual-adaptor low-rank adaptive fine-tuning architecture is adopted, and the evidence-based adapter and style adapter are trained separately on the basis of the pre-trained large language model; among them, the evidence-based adapter is fine-tuned based on evidence-based medicine guidelines and breast cancer diagnosis and treatment guidelines, and the style adapter is based on the target doctor's historical consultation records and popular science texts, extracting and transferring the target doctor's language style features from four dimensions: lexical preference, syntactic complexity, empathic expression and risk communication pattern.

[0028] S3 Boundary Layer Construction: Construct a three-level problem classification library, including a whitelist library, a blacklist library, and a gray zone problem library, and define a comprehensive confidence index:

[0029] Where C_model is the confidence level of the generated model itself, C_evidence is the confidence level of the matching based on the evidence-based medicine knowledge base, and α and β are weight coefficients. When C_total is lower than the preset threshold, a circuit breaker is triggered to prevent the content from being output to the patient.

[0030] S4 Human-Computer Collaborative Interaction: The collaborative workflow is organized in four stages: AI drafting, evidence verification, doctor review, and digital human broadcasting. In the AI ​​drafting stage, the corpus layer generates a draft response that matches the target doctor's style. In the evidence verification stage, the evidence-based knowledge base is automatically searched to trace and annotate the key medical assertions in the draft. In the doctor review stage, the verified draft is submitted to the target doctor for review and modification. In the digital human broadcasting stage, the doctor digital human generated by the image layer outputs the final approved response in the form of lip-synced video.

[0031] Step S1, which involves 3D reconstruction based on 3DGS, specifically includes: Camera parameters are estimated and sparse point clouds are generated from multi-view facial images using the Structure for Motion Restoration (SfM) algorithm. A 3D Gaussian sphere set is initialized with a sparse point cloud. Each Gaussian sphere is parameterized by the position mean μ, the covariance matrix Σ, the opacity α, and the third-order spherical harmonic coefficients, where the covariance matrix is:

[0032] The semidefiniteness is guaranteed by decomposition of the scaling matrix S and the rotation matrix R. Joint optimization is performed using L1 loss and structural similarity (SSIM) loss between the differentiable rasterization rendering results and the multi-view facial images. The loss function is:

[0033] An adaptive density control strategy is adopted to clone or split Gaussian spheres whose view space position gradient is greater than the first threshold, and to prune Gaussian spheres whose opacity is less than the second threshold.

[0034] The lip-sync model is a real-time lip-sync model based on variational autoencoder (VAE) latent space repair. After extracting audio features through Whisper encoder, it performs denoising inference in the VAE latent space through multi-scale UNet to generate a facial region image aligned with the audio. Then, the facial region image is seamlessly synthesized with the doctor's facial image rendered by 3DGS using the Poisson fusion algorithm.

[0035] Step S1 also includes a mobile compression step: The Gaussian sphere representation is quantized using INT8 to reduce storage size. Gaussian spheres that contribute less than the third threshold to the peak signal-to-noise ratio (PSNR) of the rendering are removed through structured pruning. A lightweight student network is trained to learn from the original 3DGS rendering output in a knowledge distillation manner, resulting in an image layer inference model that can run in real time on mobile devices.

[0036] The four-dimensional style features in step S2 are specifically defined as follows:

[0037] Among them, f_vocab is the vocabulary preference feature, which is represented by the difference between the word frequency-inverse document frequency (TF-IDF) vector of the target doctor's historical corpus and the TF-IDF vector of the general medical corpus; f_syntax is the syntactic complexity feature, which is represented by a weighted combination of average sentence length, clause nesting depth and dependency syntax tree complexity; f_empathy is the empathy expression feature, which is represented by the density of empathy keywords, the distribution of sentiment polarity and the ratio of first-person plural to second-person usage; f_risk is the risk communication pattern feature, which is represented by the density of uncertainty statements, the mention rate of alternatives and the frequency of disclaimer usage.

[0038] The dual-adaptor low-rank adaptive fine-tuning architecture in step S2 is specifically as follows: Let a certain weight matrix of the pre-trained large language model be... The low-rank decomposition of the evidence-based adapter is The low-rank decomposition of the style adapter is , .

[0039] The forward propagation during the reasoning phase is as follows:

[0040] Wherein λ_e and λ_s are adjustable weighting coefficients, and by controlling the contribution ratios of the evidence-based adapter and the style adapter respectively, an adjustable balance between medical accuracy and style fidelity can be achieved.

[0041] Step S3's comprehensive confidence index C_total also includes an uncertainty penalty term: when a problem in the gray zone problem library is triggered, an uncertainty penalty coefficient γ is applied to C_total, such that... , where 0 < γ < 1, and the value of γ is positively correlated with the risk level of the problem in the gray area; the risk level of the gray area is divided into three levels: low, medium and high according to the clinical risk of the problem, corresponding to γ ​​∈ [0.85, 0.95], γ ∈ [0.70, 0.85] and γ ∈ [0.50, 0.70], respectively.

[0042] Step S3 also includes incremental learning loop: When doctors revise the AI-generated draft responses during the doctor review phase, the system automatically records the text pairs before and after the revisions and incorporates these pairs into the incremental training dataset of the style adapter. When the cumulative number of valid revision pairs reaches a preset threshold, incremental fine-tuning of the style adapter is automatically triggered. During fine-tuning, the weights of the evidence-based adapter and the base weights of the pre-trained large language model are frozen, and only the low-rank matrix of the style adapter is updated. After incremental fine-tuning, the improvement in style similarity is evaluated through A / B testing, and the online style adapter is automatically replaced once the target is met.

[0043] In step S3, a three-level differentiated response strategy is adopted after the circuit breaker is triggered: Level 1 Response: When C_total∈[τ-δ,τ), output a response but with additional risk warnings and disclaimers; Level 2 response: When C_total∈[τ-2δ,τ-δ), pause automatic broadcasting and mark the issue as pending doctor review; Level 3 response: When C_total < τ-2δ, refuse to answer and output suggested information to guide the patient to seek medical treatment offline; Where τ is the total confidence threshold, δ is the buffer width, τ∈[0.65, 0.85], δ∈[0.05, 0.15].

[0044] Before constructing the corpus layer in step S2, the system also includes sensitive information desensitization processing of the target doctor's historical consultation records. The desensitization processing includes: removing or replacing patient identification information with anonymous IDs, generalizing medical institution names into institution category labels, and blurring date information to the monthly granularity, so that the training data complies with personal information protection regulations.

[0045] A digital avatar system for breast cancer doctors based on three-layer corpus engineering includes: The image layer construction module is used to perform three-dimensional reconstruction of the target doctor's multi-view facial images based on 3D Gaussian splashing, obtain a driveable Gaussian sphere representation, and realize audio-driven facial animation generation through a lip-sync model.

[0046] The corpus layer construction module is used to train the evidence-based adapter and the style adapter respectively through a dual-adaptor low-rank adaptive fine-tuning architecture, extracting and transferring the language style features of the target doctor from four dimensions: lexical preference, syntactic complexity, empathic expression and risk communication pattern.

[0047] The boundary layer control module is used to build a three-level problem classification library of whitelist, blacklist and gray zone, and to make real-time circuit breaker judgments on generated content based on a comprehensive confidence index that is a weighted fusion of model confidence and evidence confidence.

[0048] The human-computer collaborative interaction module is used to organize a collaborative workflow according to the four-stage process of AI drafting → evidence verification → doctor review → digital human broadcasting.

[0049] The core technical features of each step are summarized below: Image Layer S1: 3D reconstruction of the target doctor's multi-view facial images is performed using 3DGS. Each Gaussian sphere is parameterized by the position mean μ, covariance matrix Σ (semi-positive definiteness is ensured by decomposition using scaling matrix S and rotation matrix R), opacity α, and third-order spherical harmonic coefficients. The Gaussian sphere is projected onto the camera plane using a differentiable rasterizer renderer and optimized using the L1+SSIM joint loss function. Combined with an adaptive density control strategy (gradient-driven cloning / splitting and low-α pruning), a high-quality, driveable Gaussian sphere representation is obtained. The Gaussian sphere representation is integrated with a lip-sync model based on VAE latent space repair (such as MuseTalk) to achieve an end-to-end pipeline of audio → Whisper encoding → UNet latent space denoising → Poisson fusion → facial animation. It further includes an INT8 quantization + structured pruning + knowledge distillation compression link for mobile deployment.

[0050] Corpus S2: First Proposal of a Four-Dimensional Style Feature Space The LoRA architecture is decoupled from the dual-adaptor architecture. The evidence-based adapter A_evidence is fine-tuned using breast cancer treatment guidelines and systematic reviews such as CACA / NCCN / CSCO-BC to ensure the accuracy and traceability of medical content; the style adapter A_style learns the target physician's language style from four dimensions using historical consultation records and anonymized popular science texts. The forward propagation in the inference phase is... The dual adapters can be loaded independently, in combination, or in a weighted combination to achieve an adjustable balance between medical accuracy and style fidelity.

[0051] Boundary Layer S3: This layer introduces for the first time a three-tiered problem classification database (whitelist, blacklist, and gray zone) plus a dual-channel confidence-weighted circuit breaker mechanism. The whitelist database contains frequently asked questions related to standard breast cancer diagnosis and treatment, which can be directly answered. The blacklist database contains questions that are clearly outside the scope of practice, effectively blocking access and guiding patients to offline medical care. The gray zone database contains questions that fall between the two, triggering a comprehensive confidence calculation. C_model is calculated using the normalized value of the probability distribution entropy of the output token, and C_evidence is calculated by semantically retrieving and matching key medical assertions in the evidence-based knowledge base using BGE-M3 embedding vectors and cosine similarity. γ is a gray area penalty term. The circuit breaker response is divided into three levels: annotation reminder, suspension for review, and rejection guidance. The boundary layer also includes an incremental learning loop: automatic entry of doctor's review and modification into the database → reaching the cumulative threshold → triggering A_style incremental fine-tuning, so that the language style of the digital clone continuously approaches the doctor's current style over time.

[0052] Human-Machine Collaboration Layer S4: Connects the above three layers into a four-stage collaborative closed loop. Stage 1: AI drafting, where the corpus layer generates a draft response that matches the doctor's style. Stage 2: Evidence verification, automatically searching the evidence-based knowledge base to trace and annotate key medical assertions. Stage 3: Doctor review, submitting to the doctor for review, with modifications automatically fed back to the incremental training set of the corpus layer. Stage 4: Digital human broadcasting, generating lip-synced video output through the image layer + MuseTalk.

[0053] Compared with the prior art, the present invention has the following non-obvious beneficial effects.

[0054] 1. A unified three-in-one framework. For the first time, this approach unifies three previously fragmented dimensions—3D reconstruction of the doctor's image, 4D transfer of language style, and hierarchical circuit breaking of security boundaries—into a three-layer corpus engineering framework. These layers exhibit non-superficial synergistic benefits: the image layer provides a sense of presence for the corpus layer's output, enhancing patient trust and compliance; the corpus layer's 4D style imbues the image layer's delivery with the doctor's personal linguistic characteristics; and the boundary layer provides security guarantees for the outputs of both the image and corpus layers. These three elements are not simply superimposed but deeply coupled system integration.

[0055] 2. Dual-Adapter Decoupled Architecture. This invention introduces a dual LoRA decoupled architecture (evidence-based adapter + style adapter) for the first time in the field of doctor digital avatars. Traditional LoRA fine-tuning mixes knowledge learning and style learning in a single low-rank matrix, resulting in the inability to independently optimize medical accuracy and language personalization. This invention decouples the two into independent low-rank paths, allowing for on-demand combination during inference. During incremental learning, only the style adapter is updated while maintaining the evidence-based capability, fundamentally resolving the contradiction of "the more personalized, the less accurate." Preliminary validation shows that the dual-adaptor joint loading (λ_e=1.0, λ_s=1.0) significantly improves both medical accuracy scores (approximately 3%) and style similarity scores (approximately 9%) compared to the single-adaptor scheme.

[0056] 3. Three-tiered circuit breaker mechanism (white-black-gray). This is the first time a two-dimensional security decision matrix has been used, with the problem domain risk level as the first dimension and weighted confidence level across both channels as the second. Compared to existing solutions that rely on single-channel confidence or simple keyword filtering, this solution achieves a refined security balance: rapid whitelist approval, hard blocking of blacklists, and adaptive confidence processing in gray areas. The introduction of the gray area penalty term γ ensures the conservative security principle of "the more uncertain the problem, the more cautiously it is handled." Circuit breaker accuracy comparison experiments (see Table 2) show that, with α=0.6 and β=0.4, preliminary verification indicates that this solution is feasible in all three dimensions: whitelist approval rate, blacklist blocking rate, and correct circuit breaker rate in gray areas. Specific quantitative indicators require large-scale validation set experiments for confirmation. All three indicators are significantly better than single-channel solutions.

[0057] 4. Four-stage human-machine collaborative closed loop and incremental self-evolution. For the first time, AI drafting → evidence verification → doctor review → digital human broadcasting are linked into a four-stage integrated workflow. Moreover, the doctor's review and modification are automatically fed back as incremental training data for the style adapter. The system is no longer a static snapshot, but continuously approaches the doctor's current style as the doctor uses it. This incremental self-evolution mechanism has not been reported in existing public technologies.

[0058] 5. Cost reduction and efficiency improvement, and the downward flow of high-quality medical resources. This frees breast surgeons from a large number of repetitive standard consultations, allowing them to focus on high-value clinical decisions. Simultaneously, it provides primary healthcare institutions and patients in remote areas with a consultation experience close to that of the specialist, lowering the barrier to accessing high-quality breast specialist resources.

[0059] In the following embodiments, the experimental hardware environment consisted of a local RTX 4060 with 8GB VRAM, Windows 11, and CUDA 12.7; and a server environment with an RTX 5090 with 32GB RAM, 1 / 8 CPU (25 cores), a 90GB Xeon(R) Platinum 8470Q, a 50GB data disk, driver 580.76.05, and CUDA 13.0. The above local and server environments are merely specific implementation schemes used in this embodiment and do not constitute a limitation on the scope of protection of this invention.

[0060] Example 1: System Overall Architecture See Figure 1 The breast cancer doctor digital avatar system based on three-layer corpus engineering in this embodiment is logically divided into four layers: image layer, corpus layer, boundary layer, and human-computer collaborative interaction layer.

[0061] The image layer is responsible for the personalized appearance reconstruction and animation-driven modeling of the target doctor. Its input is a set of multi-view facial images of the doctor. (N≥15 images, covering the front and left and right sides ±45° range) and the target audio signal a(t), the output is a sequence of doctor's facial video frames synchronized with the audio. The processing flow includes: SfM sparse reconstruction → 3DGS optimization → Gaussian sphere parameter set G = {g_i} → lip-sync fusion → video frame output.

[0062] The corpus layer is responsible for injecting medical knowledge and transferring the doctor's language style. Its input consists of the patient's natural language question text q and the current conversation context c, and its output is a draft response r_draft with evidence-based source annotations and style features. The processing flow includes: question encoding → four-dimensional style feature extraction → dual-adaptor forward propagation → draft decoding and generation → evidence-based source association.

[0063] The boundary layer is responsible for the safe filtering and confidence-based circuit breaking of the generated content. Its input is the response draft r_draft generated by the corpus layer and the corresponding confidence parameters, and its output is the circuit breaking decision signal d∈{ALLOW, WARN, HOLD, BLOCK} and the corresponding patient-side output content.

[0064] The human-computer collaboration layer is responsible for the orchestration and scheduling of the four workflow segments and the closed-loop feedback of doctor review. Data and control signals are transmitted between layers through a standardized JSON interface; the interface schema definition is shown in Table 3.

[0065] Example 2: Image Layer - 3DGS Doctor's Facial Reconstruction See Figure 2Image layer section. Using the test subject as the reconstruction object, 56 multi-view facial images were collected under natural lighting conditions. The specific acquisition scheme was as follows: 1 frontal image, 2 images each from the left side at 15° / 30° / 45°, 2 images each from the right side at 15° / 30° / 45°, 1 image from a 15° downward angle, 1 image from a 15° upward angle, 1 image from a 45° downward angle from the left side, and 1 image from a 45° downward angle from the right side. The image resolution was no less than 3024×4032 pixels.

[0066] The 3D reconstruction process is as follows. First, using COLMAP's Incremental Structure from Motion (SfM) algorithm, the camera intrinsic matrix K_j, extrinsic parameters {R_j, t_j}, and sparse 3D point cloud P_sparse (approximately 11,000 3D points) are estimated from 20 images. SfM successfully registered 30 / 56 images.

[0067] The second step is to initialize a 3D Gaussian sphere set G0 = {g_i} with the sparse point cloud, where each Gaussian sphere g_i = (μ_i, Σ_i, α_i, sh_i). μ_i is initialized with the 3D coordinates of the corresponding sparse point, Σ_i is given an isotropic initial value based on the average distance of its three nearest neighbors, α_i is uniformly initialized to 0.1, and sh_i is initialized to order 0 (only the diffuse color component). The initial number of Gaussian spheres is approximately the same as the number of sparse point clouds, about 11,000.

[0068] The third step involves projecting the Gaussian sphere set onto the viewpoints of each camera using a differentiable Gaussian rasterizer to obtain the rendered image Î_j. Optimization employs a joint loss function: Where λ=0.2. L1 loss is defined as... SSIM loss is defined as The Adam optimizer was used, with an initial learning rate of 0.001 that decayed exponentially to 0.0001, for a total of 7000 iterations.

[0069] The fourth step is adaptive density control. Every 3000 iterations, a Gaussian sphere addition / removal operation is performed: (a) For Gaussian spheres with a view space position gradient greater than the threshold τ_pos=0.0002, if their world space scale is greater than τ_size=0.01 (at the normalized scene scale), then clone them; otherwise, split them; (b) remove Gaussian spheres with an opacity α less than 0.005; (c) remove anomalous Gaussian spheres with a world space scale greater than 10 times the scene radius. The number of Gaussian spheres dynamically changes during the optimization process, eventually converging to approximately 100,000-200,000.

[0070] In this embodiment, the complete reconstruction time on a single RTX 5090 is approximately 6.5 minutes, with a final number of approximately 11,000 Gaussian spheres (which can be further increased with each iteration after 7,000 iterations). The rendering PSNR is 28.03 dB, and the final L1 Loss is 0.0206. SSIM and LPIPS will be supplemented in subsequent experiments. Rendering was performed offline.

[0071] Example 3: Image Layer – Lip Synchronization Integration and Mobile Compression After completing the 3DGS reconstruction in Example 2, the lip-sync model is integrated into the image layer pipeline. This example uses the MuseTalk model, and the specific process is as follows: The first step is audio encoding. The audio signal from the doctor's response is resampled to 16kHz, and the audio feature sequence A = {a1, a2, ..., a_T} is extracted using a Whisper model (such as tiny or medium, whichever is more accurate). a}, feature dimension 512, frame rate 50Hz.

[0072] The second step is facial region VAE encoding. Facial regions (at appropriate resolution) are cropped from the current frame rendered by 3DGS and mapped to the latent space using a pre-trained VAE encoder. .

[0073] The third step is latent space conditional denoising. In the latent space, with audio feature A as a condition, random noise is iteratively denoised using a multi-scale UNet (3-level encoder-decoder structure, number of channels [64,128,256]) (DDIM sampling, 20 steps) to generate a facial latent variable z_aligned that is aligned with the audio.

[0074] The fourth step is decoding and fusion. The z_aligned image is restored to the facial region image I_face_aligned using the VAE decoder. Then, the Poisson fusion algorithm is used to seamlessly fit I_face_aligned to the corresponding region of the 3DGS rendered doctor's head image I_head to obtain the final video frame.

[0075] Lip synchronization naturalness was assessed in subjective evaluation with a subjective naturalness score of ≥4.0 / 5.0 (referencing data from the original MuseTalk paper) (5-point scale, 1 = very unnatural, 5 = very natural).

[0076] Mobile compression solution. Addressing the bottleneck of storage size and inference computation requirements of 3DGS Gaussian sphere sets, which prevent direct operation on mobile devices, this embodiment employs a three-level compression scheme.

[0077] Level 1, INT8 quantization. The Gaussian sphere parameters (μ, Σ components, α, sh coefficients) are mapped from FP32 to INT8. The quantization formula is as follows: Dequantization inverse transform: After quantization, the PSNR decreased by approximately 0.3 dB.

[0078] The second stage is structured pruning. For each Gaussian sphere, its contribution to the rendered PSNR is calculated, and they are sorted in ascending order of contribution. They are removed sequentially until the rendered PSNR drops by more than 0.5 dB or the removal rate reaches 40%. In this embodiment, the actual removal rate is approximately 32%, retaining a suitable number of Gaussian spheres (the specific number depends on the pruning threshold), allowing for further compression of storage volume.

[0079] The third stage is knowledge distillation. A lightweight student network (based on a lightweight backbone network) is trained, using the original 3DGS rendered output (FP32, full precision) as the teacher signal. The distillation loss function is... , where L_perceptual is the perceptual loss based on VGG-16.

[0080]

[0081] Table 1 Comparison of Image Layer Compression Effects Example 4: Corpus Layer - Four-Dimensional Style Feature Extraction See Figure 2 Corpus layer and Figure 3 The process for extracting four-dimensional style features is detailed below.

[0082] (1) Vocabulary preference features f_vocab. Text records of outpatient consultations and patient-oriented popular science articles from the target doctors over the past two years (a certain amount of text data after anonymization) were collected. Jieba word segmentation tool was used in conjunction with a custom breast cancer medical dictionary (containing appropriate medical terminology) for word segmentation. After removing stop words, the TF-IDF value of each word was calculated to construct the doctor's personal vocabulary vector. , where |V| is the dictionary size (appropriate size). Meanwhile, a baseline vocabulary vector V_base is constructed using a general Chinese medical corpus (MedDialog-CHS public dataset + publicly available breast cancer science articles, approximately 1.5 million Chinese characters). The top K dimensions with the largest absolute differences are used to construct the vocabulary preference feature vector. For example, a doctor might prefer to use "small lump" (TF-IDF=0.042) rather than "breast nodule" (TF-IDF=0.015), or "let's take a look" rather than "we suggest further examination." These differences are all captured by f_vocab.

[0083] (2) Syntactic complexity feature f_syntax. For each sentence in the doctor's corpus, the HanLP dependency parser was used to calculate three sub-indices: average sentence length L_avg (in words), average nesting depth of clauses D_nest, and a complexity index based on the depth of the dependency parsing tree. Where w1=0.4, w2=0.3, and w3=0.3 are empirical weights determined through ablation experiments, and _norm represents the Z-score standardized value of each indicator. f_syntax controls the syntactic complexity of the generated text. Statistically, the average sentence length and clause nesting depth of the target doctors in this embodiment vary depending on their personal style, showing a preference for a concise, short, and direct communication style that gives conclusions.

[0084] (3) Empathy expression characteristics f_empathy. An empathy keyword table was constructed (containing several Chinese medical empathy expression words / phrases, such as "understand your worries", "it is indeed not easy", "we are together", "don't be too anxious", etc.), and their occurrence density (number of occurrences per thousand characters) in the doctors' corpus was counted. At the same time, the sentiment polarity (positive / neutral / negative ratio) of each sentence in the corpus was calculated using the SnowNLP sentiment analysis tool, as well as the ratio of the use of the first-person plural ("we") to the second-person ("you / you").

[0085]

[0086] This is a 3-dimensional feature vector. The statistical values ​​of the target doctor in the above three dimensions in this embodiment are as follows: Empathy density varies depending on the doctor's personal style, which is characterized by a communication style that is mainly based on rational analysis and moderately expresses empathy.

[0087] (4) Risk communication pattern characteristics f_risk. Statistical analysis of the density of keywords expressing uncertainty, the mention rate of alternative solutions, and the frequency of use of disclaimers in the physician corpus.

[0088]

[0089] This is a 3-dimensional feature vector. The statistical values ​​of the target doctor in the above dimensions in this embodiment are as follows: the density of uncertainty expression varies depending on the doctor's personal style, which is manifested as a cautious communication style that truthfully discloses uncertainty and prefers to provide multiple solutions (the specific value varies depending on the doctor's personal style).

[0090] Example 5: Corpus Layer - Dual Adapter LoRA Fine-tuning and Comparison Experiment See Figure 3This embodiment uses Qwen2.5-1.5B-Instruct (Alibaba Tongyi Qianwen 2.5 series, 1.5 billion parameters) as the pre-trained base model. Both adapters add QLoRA low-rank decomposition (NF4 4-bit quantization, r=16, α=32) to the weights of all linear layers (including Q, K, V, O projection matrices and FFN W_up, W_down, W_gate matrices). The trainable parameters are approximately 18.46M (accounting for 2.04% of the base model), and the target layer includes all linear layers.

[0091] The choice of rank. Through ablation experiments, it was verified that the evidence-based adapter was r_evidence=16 (requires strong medical knowledge injection ability, ΔP=0.39%), and the style adapter was r_style=16 (style transfer is relatively lightweight, and excessively large r values ​​such as r=32 will lead to overfitting to the training corpus, and the style similarity of the validation set will decrease by 1.2 percentage points).

[0092] Evidence-based adapter training data. Includes: (a) the full text of the Chinese Anti-Cancer Association Breast Cancer Diagnosis and Treatment Guidelines and Standards (CACA-CBCS, 2025 edition); (b) the Chinese translation of the NCCN Breast Cancer Clinical Practice Guidelines (2025 edition); (c) the full text of the CSCO-BC Breast Cancer Diagnosis and Treatment Guidelines (2025 edition); and (d) abstracts of systematic reviews and meta-analyses in the field of breast cancer retrieved from PubMed (2020-2026, a total of 215 articles after screening). A total of 33 synthetic medical question-and-answer pairs are included. Training hyperparameter: learning_rate = 3 × 10⁻⁶ -4 batch_size=4, epochs=3, optimizer=AdamW (β1=0.9, β2=0.999, weight_decay=0.01), warmup_steps=200, cosine learning rate decayed to 1×10 -5 .

[0093] Style adapter training data. This consists of 25 evidence-based guideline paragraphs from the target physician's anonymized outpatient records and popular science articles. Training hyperparameter: learning_rate = 1 × 10⁻⁶ -4(Lower than evidence-based training to avoid overfitting), batch_size=2, epochs=5, otherwise the same as the evidence-based adapter. During training, the style similarity score on the validation set (blind scoring by 3 breast specialists, 1-5 subscale) is monitored, and the checkpoint with the highest validation set score is selected as the final model. In this example, the validation score is highest at epoch=4 (4.12 / 5.0), and the score drops to 4.05 / 5.0 at epoch=5, indicating overfitting. Therefore, the checkpoint with epoch=4 is selected.

[0094] Combinatorial strategy during the inference phase. The forward propagation formula is: The default configuration is λ_e=1.0, λ_s=1.0 (full dual adapter). When higher medical rigor is required, λ_e can be increased or λ_s can be decreased; when a closer resemblance to the doctor's natural language style is required, the settings can be reversed.

[0095] Both medical accuracy and style similarity scores used a 1-5 scale and were assessed in a blinded manner by breast surgeons. Optimal overall performance was achieved using a dual-adaptor combination.

[0096]

[0097] Table 2 Results of LoRA ablation experiments with dual adapters (mean ± standard deviation, n=50) As shown in Table 2: (a) when only the evidence-based adapter is used, medical accuracy is significantly improved, but style similarity is not significantly improved; (b) when only the style adapter is used, style similarity is significantly improved, but medical accuracy is reduced; (c) the best overall performance can be obtained by loading the dual adapters together (λ_e=λ_s=1.0), which verifies the synergistic advantages of the dual adapter decoupling architecture.

[0098] Example 6: Boundary Layer - Three-Level Problem Library and C_total Circuit Breaker See Figure 4 and Figure 5 The construction method of the three-level problem classification database is described in detail below.

[0099] Whitelist Database. This database contains 20 high-frequency standardized questions covering the entire breast cancer diagnosis and treatment cycle. Question sources: (a) High-frequency consultation records from the breast surgery outpatient clinic over the past year (questions with standard answers confirmed by doctors); (b) A list of common questions from authoritative breast cancer patient education materials. Each whitelist question is associated with 2-3 standard answer templates reviewed and confirmed by doctors, with the evidence-based source indicated (guideline name + chapter / page number). Typical whitelist questions include: "How long after surgery can the sutures be removed?" "How long should I take tamoxifen?" "Can I take a shower during radiotherapy?" "Can breast hyperplasia develop into breast cancer?" "What exercises can I do after breast cancer surgery?"

[0100] Blacklist Database. This database includes 10 questions that clearly exceed the scope of breast cancer specialist practice or pose legal risks. Questions originate from medical legal risk assessment analysis and clinical safety reviews. Typical blacklisted questions include: "Can you take a look at my lung CT report?", "How should I adjust my blood pressure medication dosage?", "I want to use only traditional Chinese medicine treatment without surgery. Do you think that's possible? (Provide definitive alternative suggestions)", and "Can your answer replace a doctor's consultation?". Blacklisted questions trigger a level-three circuit breaker response, hard-blocking the generation of a response and providing offline medical guidance information to the patient.

[0101] Gray Zone Question Database. This database contains 15 questions that fall between the whitelist and blacklist, questions that the system cannot directly apply templates to or reject. The gray zone is further subdivided into three risk levels: low-risk gray zone (e.g., lifestyle recommendations for breast cancer patients with common chronic diseases, approximately 8 questions, γ=0.90), medium-risk gray zone (e.g., discussions on individualized treatment for advanced-stage patients, approximately 5 questions, γ=0.80), and high-risk gray zone (e.g., recommendations for clinical trial participation, interdisciplinary coordination, approximately 2 questions, γ=0.70). Gray zone questions trigger C_total comprehensive confidence calculations and corresponding level circuit breaker processing.

[0102] The specific process for calculating C_total is as follows: The first step is C_model calculation. After the model generates a draft response, it calculates the information entropy of the softmax probability distribution of each token in the output sequence. , where p_i is the predicted probability of the i-th token. Normalized entropy H_norm = H / log(|V|), where |V| is the vocabulary size (the vocabulary size of Qwen2.5-1.5B-Instruct is approximately 152000). C_model ∈ [0, 1], the more confident the model is in its output (the more concentrated the probability distribution), the closer C_model is to 1.

[0103] The second step is C_evidence calculation. Key medical assertions (such as drug names and dosages, examination items and frequencies, diagnosis names, contraindications, etc.) are extracted from the generated draft responses using regular expressions and a NER model; an average of 3-8 assertions are extracted per response. For each assertion, an embedding vector is generated using the bge-small-zh-v1.5 embedding model (384 dimensions). A Top-5 nearest neighbor search is performed in the pre-built evidence-based knowledge base embedding index (FAISS IndexFlatIP index), and cosine similarity is calculated. C_evidence = the average cosine similarity of the matched assertions. Only assertions with a similarity ≥ 0.75 are considered matched; if an assertion is not matched, its contribution is considered 0.

[0104] The third step is to merge C_total. The optimal weights, determined through grid search experiments (on a validation set of 200 questions), are α=0.6 and β=0.4 (with higher weights for evidence confidence, reflecting the conservative principle of medical safety). For gray area questions, a gray area penalty is applied: , The values ​​∈ {0.90, 0.80, 0.70} correspond to three levels of gray zone risk: low, medium, and high.

[0105] Step 4: Circuit breaker determination and response. Total threshold τ = 0.70, buffer interval δ = 0.10. The three-level response strategy is shown in Table 3.

[0106]

[0107] Table 3 Three-level circuit breaker response strategy Circuit breaker accuracy comparison experiment. On the validation set (200 questions, covering 100 whitelist questions, 50 blacklist questions, and 50 graylist questions), different configuration schemes were compared and evaluated. The results are shown in Table 4.

[0108]

[0109] Table 4. Comparison of fuse failure accuracy results (n=200) Example 7: Human-Computer Collaborative Interaction - Four-Segment Closed-Loop Workflow See Figure 4 The following provides a complete description of the four-stage human-machine collaborative closed loop process and the latency data for each stage.

[0110] Phase 1 (AI drafting, end-to-end delay of approximately tens of seconds). The patient enters a text question q via WeChat mini-program or web interface → The system preprocesses q (denoising, word segmentation, entity recognition) → The boundary layer pre-classifies the question (white / black / gray) → If it is a blacklisted question, it directly jumps to the third-level response → Otherwise, the corpus layer evidence-based adapter retrieves relevant medical evidence from the evidence-based knowledge base, and the style adapter generates a draft response r_draft that matches the target doctor's language style based on the target doctor's four-dimensional style characteristics → Each key medical assertion in the draft is labeled with its evidence-based source.

[0111] Phase Two (Evidence Verification, delayed by approximately several seconds). The system automatically extracts the key medical assertion set E = {e1, ..., e_m} from r_draft → uses BGE-M3 vector search to retrieve the Top-5 evidence-based knowledge bases for each e_i → marks the source information of matched assertions (guideline name + chapter number + page number or paragraph number) → highlights unmatched assertions as "to be verified" → generates a verification report, including the traceability score of each assertion.

[0112] Phase Three (Doctor Review, manual review, typically a few minutes). The verified draft and verification report are pushed to the doctor's app / web interface → Doctor views the draft and verification report → Optional actions: (a) One-click confirmation for publication; (b) Publish after modifications; (c) Reject and manually rewrite. The doctor's modifications are automatically recorded by the system: saving "text before modification - text after modification" pairs (diff format), and indicating the modification type (content correction / style adjustment / supplementary information / redundancy removal). Each modification record includes a timestamp and the anonymized context of the original question that triggered the response.

[0113] Phase Four (Digital Human Broadcast, delayed by approximately tens of seconds). The final response text r_final, which has passed review, is sent to the MuseTalk pipeline in the image layer, generating a lip-synced video of the target doctor, adding subtitles and evidence-based source annotations, and pushing it to the patient's end, where the patient watches the doctor's digital human video response.

[0114] Incremental learning closed loop. The "before-modification - after-modification" text pairs accumulated in Phase 3 are automatically added to the incremental training dataset D_incremental. → When the cumulative number of valid modification pairs exceeds a preset threshold, → incremental fine-tuning of the style adapter A_style is automatically triggered (ΔW_evidence and base weights W are frozen during training, only ΔW_style is updated). → After fine-tuning, A / B testing (comparing the style similarity scores of the old and new style adapters on the same 50 test questions) is automatically run on the reserved evaluation set. → If the score improvement reaches a preset standard, the online style adapter is automatically replaced; otherwise, the old version is retained and the operations team is notified for manual decision-making. This closed loop ensures that the language style of the digital clone continuously approaches the target doctor's current style over time.

[0115] Example 8: End-to-end scenario The following is a complete end-to-end use case to illustrate the synergistic effect of the various layers of this invention in practical applications.

[0116] Scenario: Ms. Zhang, 45 years old, is three weeks post-modified radical mastectomy for breast cancer. At 21:15 in the evening, she initiates a consultation with the digital avatar of her attending physician (Dr. Wang, chief physician of breast surgery) via WeChat mini-program.

[0117] The patient asked: "Dr. Wang, it's been 3 weeks since my surgery, and the drainage tube has been removed, but the area around the wound has become a bit red and itchy in the last few days. Is this normal? Should I come to the hospital to get it checked out?" The system processing flow is as follows: (1) Question pre-classification: The boundary layer pre-classifies the questions. The main classification is determined to be the whitelist framework (consultation on common symptoms during the postoperative wound healing period), but the question "slight redness" triggers the gray area sub-condition (it is necessary to distinguish between normal healing and infection). The system recognizes it as the superposition of whitelist and gray area sub-conditions, triggering the C_total comprehensive judgment.

[0118] (2) C_total calculation: C_model shows that the output probability distribution of the model for postoperative wound care issues is relatively concentrated; C_evidence shows that the content related to postoperative wound healing care in the evidence-based library has a high degree of matching, with 2 out of 3 key assertions achieving a Top-1 similarity of >0.80; C_total is calculated by weighted fusion. Because it is within the whitelist framework and the gray zone risk level is judged to be low (γ=0.90), C_total′ is penalized by applying gray zone penalty. C_total′ falls into the secondary response interval, triggering the secondary response. After generating a response, the automatic broadcast is paused and pushed to the doctor's review end.

[0119] (3) AI drafting: The corpus layer generates the response draft. The draft reflects Dr. Wang's four-dimensional style characteristics, using "we" to establish a sense of alliance (f_empathy), "relatively common" and "under normal circumstances" to reflect a moderate degree of uncertainty (f_risk), a hierarchical structure that explains the pathological mechanism before giving suggestions (f_syntax), and patient-friendly terms such as "redness" and "itching" (f_vocab). The draft contains three evidence-based sources: "Neovascularization and tissue remodeling may cause local redness and itching" → CACA-CBCS postoperative management chapter (section 4.2); "Be alert to the possibility of incision infection" → NCCNSurvivorship Guideline 2025 (SURV-8); "No fever, no purulent discharge, no expansion of redness and swelling → can be observed first" → UpToDate: Postoperative wound care (2025).

[0120] (4) Doctor's review: At 21:17, Dr. Wang received the review notification via the mobile app. After reviewing the draft and verification report, Dr. Wang changed the last part, "If the above situation occurs, please come to the hospital for a follow-up examination," to "If the above situation occurs, please come to my clinic for a follow-up examination on Tuesday morning," and clicked "Confirm and publish." This modification was automatically recorded by the system as a "content correction" type and included in the incremental training dataset.

[0121] (5) Digital Human Broadcast: The revised final text is sent to the MuseTalk pipeline in the image layer to generate a lip-sync video of Dr. Wang's face, with added subtitles and 3 evidence-based source annotations. At 21:20, the video is pushed to the patient Ms. Zhang's WeChat.

[0122] The entire process, from the patient asking a question to receiving a digital human video response, takes about 5 minutes (of which the system automatically processes the information for about tens of seconds, the doctor reviews it manually for about a few minutes, and the network transmission delay is negligible). Compared with the traditional offline clinic model (where patients need to register, queue, have a face-to-face consultation, and communicate repeatedly, the total time usually takes ≥2 hours), efficiency is significantly improved.

[0123] Example 9: Deployment Architecture and Privacy Protection This embodiment describes the system's deployment architecture and privacy protection measures in a production environment.

[0124] The system adopts a hybrid deployment architecture of cloud + edge. Offline training tasks (3DGS reconstruction, LoRA evidence-based / style adapter training, knowledge distillation student network training) are completed on the cloud GPU cluster (configuration: multi-GPU cluster), and the training products (Gaussian sphere parameter file, LoRA weight matrix, student network weights) are compressed and signed before being deployed to the edge inference server.

[0125] The online inference pipeline (Whisper encoding + MuseTalk latent space inference + Qwen3-8B LoRA inference + C_total computation + FAISS vector retrieval) all runs on edge inference servers to reduce end-to-end latency (target <100ms) and meet the localized storage requirements of medical data. The edge servers are deployed within the hospital's internal network and isolated from the internet via firewalls and VPNs.

[0126] Patients interact with the edge server via a WeChat mini-program or web app (HTTPS / TLS 1.3 encryption). Doctors receive review notifications via a dedicated app (VPN access to the hospital's intranet). All patient questions (text) and responses (video) are stored with AES-256-GCM encryption, and database access uses role-based access control (RBAC).

[0127] The privacy protection processing of the training data includes: (a) patient name → anonymous ID (SHA-256 hash + salt value); (b) mobile phone number / ID number / address → physical deletion; (c) specific name of medical institution → institution category label (e.g., "Breast Surgery Department of a Tertiary Hospital"); (d) date of visit → month granularity (e.g., "2025-03-15" → "2025-03"). After the above desensitization processing, the training data meets the compliance requirements of the "Personal Information Protection Law of the People's Republic of China" (2021) and the "Interim Measures for the Security Management of Health and Medical Big Data".

[0128] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing digital avatars for breast cancer doctors based on three-layer corpus engineering, characterized in that, Image layer construction: Acquire multi-view facial images of the target doctor, perform three-dimensional reconstruction based on 3D Gaussian splashing, obtain a driveable Gaussian sphere representation, integrate the Gaussian sphere representation with the lip-sync model, and realize audio-driven facial animation generation. Corpus layer construction: A dual-adaptor low-rank adaptive fine-tuning architecture is adopted. The evidence-based adapter and the style adapter are trained separately on the pre-trained large language model. During the inference stage, the contribution ratio of the evidence-based adapter and the style adapter is controlled by weighting coefficients. Boundary layer construction: Construct a whitelist library, a blacklist library, and a gray zone problem library. Define a comprehensive confidence index that is a weighted fusion of the model's own confidence and the confidence of evidence-based evidence. When the comprehensive confidence index is lower than a preset threshold, a circuit breaker is triggered. When a gray zone problem is triggered, an uncertainty penalty coefficient is applied to the comprehensive confidence index.

2. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, This includes extracting language style features from four dimensions: lexical preferences, syntactic complexity, empathic expression, and risk communication patterns.

3. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, The uncertainty penalty coefficient is mapped to three levels of gray zone risk level. The uncertainty penalty coefficient for low risk level is 0.85 to 0.95, for medium risk level it is 0.70 to 0.85, and for high risk level it is 0.50 to 0.

70.

4. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, In the comprehensive confidence index, the model's own confidence is calculated based on the normalized information entropy of the probability distribution of the output sequence, while the evidence-based confidence is calculated based on the semantic retrieval matching degree between key medical assertions in the generated content and the evidence-based knowledge base.

5. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, This includes obtaining the target doctor's modifications to the draft response, recording the text pairs before and after the modifications, and triggering incremental fine-tuning of the style adapter when the cumulative number of valid modification pairs reaches a preset threshold. During fine-tuning, the weights of the evidence-based adapter and the base weights of the pre-trained large language model are frozen, and only the low-rank matrix of the style adapter is updated.

6. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, This includes INT8 quantization of the Gaussian sphere representation to reduce storage size, removing Gaussian spheres whose contribution to the peak signal-to-noise ratio of the rendering is lower than a preset contribution threshold through structured pruning, training a lightweight student network, and learning from the original 3D Gaussian splash rendering output through knowledge distillation to obtain an image layer inference model that can run in real time on mobile devices.

7. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 5, characterized in that, After incremental fine-tuning is completed, A / B testing is used to evaluate the style similarity before and after the style adapter is updated. Once the style similarity improvement reaches the preset standard, the online style adapter is automatically replaced.

8. The method for constructing a digital avatar for breast cancer doctors based on three-layer corpus engineering according to claim 1, characterized in that, The process involves organizing a collaborative workflow according to four stages: AI drafting, evidence verification, doctor review, and digital human broadcasting. In the AI ​​drafting stage, a style adapter generates a draft response that matches the target doctor's style. In the evidence verification stage, the system automatically searches the evidence-based knowledge base to trace and annotate key medical assertions in the draft. In the doctor review stage, the verified draft is submitted to the target doctor for review and modification. In the digital human broadcasting stage, the final approved response is output in the form of lip-synced video.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing instructions, which, when executed by the processor, implement the method for constructing a digital avatar of a breast cancer doctor based on three-layer corpus engineering as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for constructing a digital avatar of a breast cancer doctor based on three-layer corpus engineering as described in any one of claims 1 to 8.

Images