Medical vision-language model prejudice relieving method and system based on Logit adjustment
By adjusting the framework's adaptive selection bias correction strategy using Logit, the bias mitigation problem of medical vision-language models in different task scenarios is solved, improving the fairness and reliability of the model and making it suitable for lightweight deployment of large closed-source models.
Patent Information
- Application Number
- CN202511314955.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-15
- Publication Date
- 2025-12-16
AI Technical Summary
Existing medical vision-language models suffer from insufficient cross-task consistency and insufficient sample-level adaptation in mitigating bias, resulting in limited mitigation effects across different task scenarios. Furthermore, existing methods are computationally expensive and difficult to deploy on large-scale closed-source models.
By adopting the Logit adjustment framework, the bias boundary is defined by calculating the Logit difference between valid and invalid inputs. The Logit smoothing or bias compensation strategy is adaptively selected to achieve dynamic bias correction and improve the applicability and robustness of the model in different task scenarios.
It achieves unified bias correction in classification and generation tasks, improves the fairness and reliability of the model, and eliminates the need to retrain the model. It is also lightweight and can be embedded into existing inference processes, ensuring deployment efficiency.
Smart Images
Figure CN121148655A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of natural language processing technology, and in particular relates to a method and system for mitigating bias in medical vision-language models based on Logit adjustment. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Large-scale medical vision-language models have demonstrated outstanding performance in tasks such as clinical diagnosis and medical reasoning. However, with the increase in model size and capabilities, the amplification effect of inherent biases in pre-training data becomes increasingly prominent. Such biases not only undermine the fairness and credibility of the model but may also pose potential risks to clinical decision support. Therefore, how to effectively mitigate and correct the bias problem in medical vision-language models has become a crucial technical challenge that urgently needs to be addressed.
[0004] Based on the intervention stage of bias mitigation methods, existing methods can be divided into three categories: preprocessing, in-processing, and postprocessing. Data balancing and counterfactual augmentation are common preprocessing methods that reduce bias by adjusting the input distribution or replacing sensitive attributes, eliminating or weakening bias signals at the data level. Their limitation lies in their reliance on high-quality original pre-training data and fine annotations, which are difficult to obtain in large medical models, and they lack robustness in cases of clinical distribution shifts. In-processing methods, including adversarial debiasing, training data constraints, and robust representation learning, can directly suppress bias during model training and are suitable for scenarios requiring deep control over model parameters. However, these methods often require retraining the entire model, incurring huge computational costs, making them difficult to apply to large-scale and closed-source medical models. Postprocessing methods, by calibrating the model output during the inference stage, have advantages such as being model-independent, requiring no parameter modification, and being easy to deploy, thus offering greater flexibility in practical applications.
[0005] However, existing post-processing methods have two major shortcomings: First, there is a lack of a unified processing mechanism across tasks (such as classification and open generation); second, the adjustment strategy is fixed and cannot be adaptively corrected according to the degree of bias in the samples, resulting in limited mitigation effect. Summary of the Invention
[0006] To address at least one of the technical problems in the background art, the present invention provides a method and system for mitigating bias in medical vision-language models based on Logit adjustment. This method and system can perform dynamic quantification and correction post-processing based on the bias intensity of different samples, thereby significantly improving the fairness and reliability of large-scale medical vision-language models.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for mitigating bias in a medical vision-language model based on Logit adjustment, comprising the following steps: Obtain image information and text query information; The image information and text query information are used as inputs to the medical vision-language model; Medical vision-language models generate predictive responses based on conditional probability distributions; Define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution; Based on the expected Logit, calculate the biased Logit distance; A quantization mechanism based on bias boundaries is introduced, and a bias correction strategy is adaptively selected based on the debiased Logit distance to obtain the corrected prediction distribution.
[0008] As a further implementation scheme, the medical vision-language model generates predicted responses based on conditional probability distributions, specifically: (1) in, For image information, The query information is text, and θ represents the model parameters.
[0009] As a further implementation plan, the medical vision-language model is defined for downstream task distribution. With training distribution The expected Logit on is respectively and : (2).
[0010] As a further implementation scheme, the method for calculating the biased Logit distance is as follows: (3) in, Image information indicating invalid input. Image information that is valid input.
[0011] As a further implementation scheme, the introduction of a quantification mechanism based on bias boundaries specifically involves: Let The Logit distance threshold between valid and invalid inputs is determined by... The size of the selection strategy is adaptive.
[0012] As a further implementation plan, when When the value reaches a certain threshold, it indicates a small bias boundary, suggesting that the model over-relies on artifacts. Logit smoothing is then employed. (4) By scaling and fusing the Logits of valid and invalid inputs, the model's overconfidence is reduced, thereby avoiding biased outputs. (0, 0.5) is used to scale the confidence level of the model; when This indicates a large bias boundary, meaning the valid input contains important medical information, and bias compensation is employed. (5) in, This is the bias correction coefficient.
[0013] A second aspect of the present invention provides a Logit-adjusted medical vision-language model bias mitigation system, employing the steps of the method provided in the first aspect, including: The information acquisition module is configured to acquire image information and text query information; The information input module is configured to take the image information and text query information as input to the medical vision-language model; The data processing module is configured to: define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution; calculate the bias-reduced Logit distance based on the expected Logit; introduce a quantization mechanism based on the bias boundary; and adaptively select a bias correction strategy according to the bias-reduced Logit distance to obtain the corrected prediction distribution. The dynamic post-processing output module is configured to output the corrected predicted distribution.
[0014] A third aspect of the present invention provides a computer-readable storage medium.
[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect described above.
[0016] A fourth aspect of the present invention provides a computer program product.
[0017] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method provided in the first aspect described above.
[0018] A fifth aspect of the present invention provides an electronic device.
[0019] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method provided in the first aspect above.
[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. To address the lack of a unified cross-task mechanism in existing post-processing methods, this invention constructs a general Logit adjustment framework. This framework enables consistency in bias correction strategies across classification and open-ended generation tasks by performing unified operations on Logit during the inference phase, thereby improving the applicability and robustness of the method in different medical application scenarios.
[0021] 2. To address the problem that existing methods generally employ fixed correction strategies and struggle to dynamically adjust based on sample differences, this invention proposes a quantification mechanism based on bias boundaries. Specifically, it first calculates the Logit difference between valid and invalid inputs to characterize the bias intensity of a single sample; then, based on the magnitude of the bias intensity, it adaptively selects either Logit smoothing or bias compensation strategies to achieve dynamic correction at the sample level, thereby effectively avoiding under-correction or over-correction.
[0022] 3. The method of the present invention does not require retraining the model or accessing pre-trained data, and can be embedded into the inference process of existing medical vision-language models in a lightweight manner, which not only ensures deployment efficiency, but also enhances the fairness of the model and the reliability of diagnostic prediction.
[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0025] Figure 1 This is a flowchart of the proposed method for mitigating bias in a medical vision-language model based on Logit adjustment. Figure 2 This is a flowchart of the scheme for the Logit-based bias mitigation system for medical vision-language models in this invention; Figure 3 This is a framework diagram of the adaptive selection bias correction strategy of this invention. Detailed Implementation
[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0029] Example 1 This invention addresses the shortcomings of bias mitigation methods in large-scale medical vision-language models, focusing on solving the following two core problems: (1) Insufficient cross-task consistency: Existing bias mitigation methods are usually designed for single tasks and lack a unified processing mechanism across tasks, making it difficult to maintain consistent bias mitigation effects in various clinical application scenarios such as classification and generation. This invention proposes a generalized Logit adjustment framework that can achieve a unified bias correction strategy under different task types, thereby improving the applicability and robustness of the method.
[0030] (2) Insufficient sample-level adaptation: Most existing methods adopt fixed adjustment strategies, which cannot dynamically correct according to the degree of bias of specific samples, easily leading to undercorrection or overcorrection of some samples. This invention proposes a bias intensity quantification mechanism based on Logit difference, and implements sample-level dynamic smoothing and compensation accordingly, thereby achieving more accurate bias mitigation and effectively improving the fairness and reliability of medical vision-language models in clinical applications.
[0031] The embodiment provides a method for mitigating bias in a medical vision-language model based on Logit adjustment, such as Figure 1 As shown, it includes the following steps: Step 1: Obtain image information and text query information.
[0032] The image information is a random medical image, and the text query information is a medical question related to that medical image.
[0033] Step 2: Use the image information and text query information as input to the medical vision-language model.
[0034] Among them, medical vision-language models are existing large models, such as LLaVA-Med and Qwen-VL-7B.
[0035] The medical vision-language model generates predicted responses based on conditional probability distributions, specifically: (1) in, For image information, Here, θ represents the text query information, and θ represents the model parameters. Ideally, the predicted distribution should depend on the input medical evidence; however, in practice, biases present in the pre-training corpus can cause statistical artifacts to affect the model output.
[0036] Step 3: Define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution.
[0037] To reveal the perpetuation and amplification effect of pre-training bias in downstream tasks, we define the model distribution in downstream tasks. With training distribution The expected Logit on is respectively and : (2) If the above relationship holds, it indicates that even after the introduction of new evidence, the model predictions are still significantly constrained by the pre-training priors, i.e., there is a catastrophic inheritance of bias.
[0038] Step 4: Calculate the biased Logit distance based on the expected Logit.
[0039] like Figure 2 As shown, invalid inputs (such as completely black, completely white, or noisy images) are introduced as a reference, and the Logit difference between the valid and invalid inputs is calculated, defined as the biased Logit distance: (3) in, Image information indicating invalid input. This is valid input image information. If... The lower the value, the more the model relies on pre-training priors; if the value is higher, the model relies more on input evidence.
[0040] Step 5: Introduce a quantification mechanism based on bias boundaries, and adaptively select a bias correction strategy based on the debiased Logit distance.
[0041] make The Logit distance threshold between valid and invalid inputs is determined by... The size is determined by adaptively choosing one of the following two strategies: (1) When When the value reaches a certain threshold, it indicates a small bias boundary, suggesting that the model over-relies on artifacts. Logit smoothing is then employed. (4) By scaling and fusing the Logits of valid and invalid inputs, the model's overconfidence is reduced, thereby avoiding biased outputs. (0, 0.5) is used to scale the confidence level of the model.
[0042] (2) When This indicates a large bias boundary, meaning the valid input contains important medical information, and bias compensation is employed. (5) in, This is the bias correction coefficient, used to amplify the contribution of effective inputs and suppress interference from invalid inputs.
[0043] By introducing a quantization mechanism based on bias boundaries, adaptive adjustments can be made according to the intensity of sample bias, thereby avoiding under-correction or over-correction.
[0044] Step 6: Dynamic Post-processing Output. During the inference phase, based on the detected degree of bias, either Logit smoothing or bias compensation strategy is dynamically selected to obtain the corrected prediction distribution: (6) in, This is the softmax normalization function. Dynamic post-processing ensures the consistency of bias mitigation methods across multi-task scenarios and achieves more accurate fairness enhancement at the sample level.
[0045] Example 2 This embodiment provides a bias mitigation system for a medical vision-language model based on Logit adjustment, such as... Figure 2 As shown, the steps of the Logit-based bias mitigation method for medical vision-language models provided in Example 1 include: The information acquisition module is configured to acquire image information and text query information; The information input module is configured to take image information and text query information as input to the medical vision-language model; The data processing module is configured to: define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution; calculate the bias-reduced Logit distance based on the expected Logit; introduce a quantization mechanism based on the bias boundary; and adaptively select a bias correction strategy according to the bias-reduced Logit distance to obtain the corrected prediction distribution. The dynamic post-processing output module is configured to output the corrected predicted distribution.
[0046] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the Logit-based bias mitigation method for medical vision-language models as described above.
[0047] Example 4 A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the Logit-adjusted medical vision-language model bias mitigation method as described above.
[0048] Example 5 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the Logit-based bias mitigation method for medical vision-language models described above.
[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for mitigating bias in a medical vision-language model based on Logit adjustment, characterized in that, Includes the following steps: Obtain image information and text query information; The image information and text query information are used as inputs to the medical vision-language model; Medical vision-language models generate predictive responses based on conditional probability distributions; Define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution; Based on the expected Logit, calculate the biased Logit distance; A quantization mechanism based on bias boundaries is introduced, and a bias correction strategy is adaptively selected based on the debiased Logit distance to obtain the corrected prediction distribution.
2. The method for mitigating bias in medical vision-language models based on Logit adjustment as described in claim 1, characterized in that, The medical vision-language model generates predicted responses based on conditional probability distributions, specifically: （1） in, For image information, The query information is text, and θ represents the model parameters.
3. The method for mitigating bias in medical vision-language models based on Logit adjustment as described in claim 1, characterized in that, Defining the medical vision-language model in downstream task distribution With training distribution The expected Logit on is respectively and : （2）。 4. The method for mitigating bias in medical vision-language models based on Logit adjustment as described in claim 3, characterized in that, The method for calculating the bias correction Logit distance is as follows: （3） in, Image information indicating invalid input. Image information that is valid input.
5. The method for mitigating bias in medical vision-language models based on Logit adjustment as described in claim 4, characterized in that, The introduction of a quantization mechanism based on bias boundaries specifically involves: Let The Logit distance threshold between valid and invalid inputs is determined by... The size of the selection strategy is adaptive.
6. The method for mitigating bias in medical vision-language models based on Logit adjustment as described in claim 5, characterized in that, when When the value reaches a certain threshold, it indicates a small bias boundary, suggesting that the model over-relies on artifacts. Logit smoothing is then employed. （4） By scaling and fusing the Logits of valid and invalid inputs, the model's overconfidence is reduced, thereby avoiding biased outputs. (0, 0.5) is used to scale the confidence level of the model; when This indicates a large bias boundary, meaning the valid input contains important medical information, and bias compensation is employed. （5） in, This is the bias correction coefficient.
7. A medical vision-language model bias mitigation system based on Logit adjustment, employing the medical vision-language model bias mitigation method based on Logit adjustment as described in any one of claims 1-6, characterized in that, include: The information acquisition module is configured to acquire image information and text query information; The information input module is configured to: construct a medical vision-language model and use the image information and text query information as input to the medical vision-language model; The data processing module is configured to: define the expected Logit of the medical vision-language model on the downstream task distribution and training distribution; calculate the bias-reduced Logit distance based on the expected Logit; introduce a quantization mechanism based on the bias boundary; and adaptively select a bias correction strategy according to the bias-reduced Logit distance to obtain the corrected prediction distribution. The dynamic post-processing output module is configured to output the corrected predicted distribution.
8. 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 steps in the Logit-adjusted bias mitigation method for medical vision-language models as described in any one of claims 1-6.
9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the Logit-adjusted medical vision-language model bias mitigation method as described in any one of claims 1-6.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the Logit-adjusted medical vision-language model bias mitigation method as described in any one of claims 1-6.