Chest radiograph report generation method and system based on thinking chain reasoning and active learning
By employing a multi-perspective fusion and structured five-stage chain reasoning method for generating chest X-ray reports, combined with a scoring function for uncertainty and inconsistency, this approach addresses the issues of disordered diagnostic processes and high annotation costs in existing technologies, achieving efficient and interpretable chest X-ray report generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-13
AI Technical Summary
Existing methods for generating chest X-ray reports lack a structured diagnostic process consistent with clinical practice, have weak reasoning capabilities, and require high-quality data annotation at a high cost, which limits the efficiency of model training and iteration.
By employing chain reasoning and active learning methods, and through multi-perspective fusion and structured five-stage chain reasoning, combined with a scoring function for uncertainty and inconsistency, a systematic hierarchical diagnostic process is constructed to generate structured chest X-ray reports.
It significantly improves the interpretability and accuracy of reports, reduces annotation and training costs, enhances the robustness and usability of models, and adapts to annotation environments with limited resources.
Smart Images
Figure CN121661397A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chest X-ray image analysis and chest X-ray report generation technology, and in particular to a chest X-ray report generation method based on thought chain reasoning and active learning. Background Technology
[0002] Automatic chest X-ray report generation aims to automatically generate reports based on input chest X-ray images to assist radiologists in report writing. With the rapid development of multimodal large language models, their integration into the report generation process has become a mainstream trend. However, existing methods still have the following shortcomings: 1) Lack of a structured diagnostic process consistent with clinical practice. Most methods directly generate text end-to-end from overall image features, without explicitly following the reasoning sequence of "perspective-anatomical structure-diagnostic steps," resulting in weak reasoning ability. 2) Existing methods lack a systematic diagnostic reasoning process that conforms to clinical norms; most directly generate reports end-to-end, failing to fully reflect the physician's detailed diagnostic steps for each anatomical structure. 3) High cost of high-quality reasoning data annotation. Constructing clinically usable data annotations requires repeated proofreading and revision by professional physicians, resulting in significant costs and time commitments, limiting the training and iteration efficiency of the model. Summary of the Invention
[0003] In view of the aforementioned shortcomings of existing technologies, this invention proposes a multi-perspective chest X-ray report generation method based on thought chain reasoning and active learning. It employs thought chain reasoning to structurally model the doctor's diagnostic thought process, performing detailed reasoning on each anatomical structure of the chest X-ray, thereby improving the reliability and interpretability of the generated report. Through an active learning-based sample selection strategy, the accuracy, interpretability, and robustness of the generated report are improved while controlling annotation costs.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0005] The first aspect is a chest X-ray report generation method based on thought chain reasoning and active learning, which includes the following steps:
[0006] S1. Acquire and preprocess multi-view chest X-ray images and their corresponding reports; perform viewpoint identification and viewpoint classification on the multi-view chest X-ray images;
[0007] S2. A shared visual encoder is used to extract sequence features from each viewpoint, and multi-view visual tokens are generated through cross-viewpoint feature fusion.
[0008] S3. Construct a systematic hierarchical diagnostic process, and sequentially execute structured five-stage chain reasoning for each anatomical structure in multi-view chest X-ray images to generate structured reasoning prompts.
[0009] S4. Input the multi-view visual token and the structured reasoning prompt into the large language model, and generate the target sequence through autoregression.
[0010] S5. Construct a scoring function based on the joint scoring of uncertainty and inconsistency; in each iteration, label the samples in the unlabeled sample pool according to the sample scores to train the language model for generating chest X-ray reports;
[0011] S6. Input the multi-view chest X-ray images to be tested into the trained language model to generate the thought chain and complete report text.
[0012] Preferably, in S2, an Encoder(·) visual encoder is used to extract the sequence features for the i-th single-view image I. i :
[0013]
[0014] Among them, f i Let M be the sequence features of the i-th single-view image, and M be the sequence length.
[0015] As a preferred approach, in S2, centralized multi-head attention is used for cross-perspective fusion and residual convergence:
[0016]
[0017] V = {v1, v2, ..., v} M}∈R M×d
[0018] Where N is the number of viewpoints, M is the sequence length, d is the feature dimension, V is the multi-view visual token, and MHA(·) represents the multi-head attention mechanism.
[0019] Preferably, to achieve detailed reasoning about each anatomical structure, the structured intermediate reasoning in S3 includes, in sequence:
[0020] Region localization: Fine-grained region division and localization of the target structure;
[0021] Disease identification: Identify potential pathologies or abnormalities and provide high-level interpretations;
[0022] Severity assessment: The degree of lesion is characterized using a discrete scale;
[0023] Probabilistic reasoning: Based on contextual clues, provide a probabilistic analysis of comorbidities, interrelated abnormalities, or potential causes;
[0024] Evidence-based interpretation and summary: Summarize the aforementioned information and output a structured summary statement.
[0025] As a preferred embodiment, S4 also includes generating the inference chain generation probability through an autoregressive approach, modeled as follows:
[0026]
[0027] Among them, c t C is the thought chain token output at the current moment. 0:t-1 Let x = concat{V, prompt}, where prompt is the structured reasoning instruction generated from the chest X-ray report of the input large language model, and T is the length of the thought chain.
[0028] Preferably, the training objective of the language model is cross-entropy loss, and the cross-entropy loss formula is constructed based on the inference chain generation probability as follows:
[0029]
[0030] Preferably, the scoring function described in S5 is:
[0031]
[0032] in, For uncertainty, i.e., the entropy average based on the top-1 probability of the generated sequence, the calculation method is as follows:
[0033]
[0034] Where, p t =p(c t |C 0:t-1 x) represents the top-1 probability of generating the t-th token. Inconsistency is indicated by a consistency score based on the degree of overlap of medical entities, calculated as follows:
[0035]
[0036]
[0037] Among them, E c With E r These are the sets of medical entities extracted from the generated reasoning report and the reference report, respectively.
[0038] Secondly, a chest X-ray report generation system based on thought chain reasoning and active learning includes:
[0039] The perspective recognition module is used to acquire multi-view chest X-ray images and their corresponding reports, perform perspective recognition on the multi-view chest X-ray images, and generate single-view images.
[0040] A visual encoding module is used to extract sequential features from each of the single-view images using a shared visual encoder;
[0041] A cross-view fusion module is used to fuse the sequence features to generate a multi-view visual token set;
[0042] The structured reasoning module is used to construct a hierarchical diagnostic process that first determines the perspective and then systematically examines each anatomical structure from that perspective. For each anatomical structure, it performs structured five-stage chain reasoning to generate a structured intermediate reasoning chain.
[0043] The reasoning and decoding module is used to generate target sequences based on a multi-view visual token set and structured reasoning hints, using an autoregressive approach.
[0044] The active learning selection module is used to construct a scoring function based on the joint scoring of uncertainty and inconsistency, and to label samples in the unlabeled sample pool according to the sample scores;
[0045] The training module is used to perform cross-entropy loss optimization and active learning iteration;
[0046] The chest X-ray report generation system based on thought chain reasoning and active learning is used to implement the chest X-ray report generation method based on thought chain reasoning and active learning as described in the first aspect.
[0047] Compared with the prior art, the beneficial effects of the present invention are reflected in:
[0048] 1. Unlike traditional technologies that often generate reports from a single perspective or in an unordered end-to-end manner, this invention adopts a multi-perspective fusion and structured five-stage reasoning strategy to achieve a reasoning process that proceeds in the order of diagnosis based on perspective and anatomical structure. This significantly improves the completeness and rationality of the report content and enhances the explanatory power and application value of the model.
[0049] 2. Unlike existing active learning methods that rely on a single uncertainty measure, this invention combines uncertainty based on generated sequence entropy with inconsistency scoring of medical entity overlap, which strengthens the semantic relevance of sample selection, improves the effectiveness of sample selection and the utilization rate of annotation resources, reduces training costs and maintains superior model performance.
[0050] 3. This invention achieves deep fusion of information between perspectives through a multi-view visual encoder and a centralized multi-head attention mechanism, making full use of complementary information from images in different orientations, enhancing the ability to perceive diagnostic context, and improving the difficulty of interpreting multi-view chest X-rays.
[0051] In summary, the proposed technical solution effectively reduces the difficulty of annotation and training while ensuring model performance, and improves the interpretability, reliability, accuracy, robustness and practicality of automatically generated chest X-ray reports, showing promising prospects for widespread application. Attached Figure Description
[0052] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention;
[0053] Figure 2 This is a schematic diagram of the hierarchical report generation process of Embodiment 1 of the present invention;
[0054] Figure 3 This is a schematic diagram of the active learning sample selection process in Embodiment 1 of the present invention;
[0055] Figure 4 This is a comparative example of the reports generated by MAIRA-2, LLaVa-Rad, MedGemma, and the CACoT method of this invention in Embodiment 5 of the present invention;
[0056] Figure 5 This is a diagram illustrating the ablation effect of Embodiment 5 of the present invention. Detailed Implementation
[0057] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.
[0058] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.
[0059] Specifically, this invention employs a large language model as a text generation decoder, fusing multi-view chest X-ray image features as visual cues for the large language model, and inputting text instructions to generate reports, generating thought chain reasoning for each anatomical structure according to the diagnostic sequence. The thought chain reasoning in this invention innovatively constructs a clinically aligned hierarchical diagnostic process, refining it into a five-stage reasoning chain (regional localization, disease identification, severity assessment, probabilistic reasoning, and evidence-based summary) based on the perspective information and anatomical structure sequence of multi-view images. This effectively enhances the model's systematic understanding of complex lesions on chest X-rays, overcoming the problems of disordered, lengthy, and unrelated thought chain reasoning in existing technologies, and improving the accuracy and interpretability of the reasoning. Subsequently, the thought chains of all chest X-ray regions are integrated to obtain the chest X-ray report. Secondly, to construct training data labeled with thought chains, this invention introduces an active learning strategy, continuously using the previously trained model to predict the most valuable samples in new samples, and labeling selected samples with thought chains, incorporating them into the next round of model training. The active learning strategy proposed in this invention innovatively combines a dual scoring mechanism based on the predictive uncertainty of generated sequences and the inconsistency of medical entity overlap, effectively identifying samples that contribute most to model improvement. This significantly reduces the manual cost of high-quality thought chain annotation and improves model training efficiency and diagnostic performance, making it particularly suitable for annotation environments with limited resources. Quantitative and qualitative experiments on public datasets demonstrate that this method achieves state-of-the-art performance.
[0060] Example 1:
[0061] like Figure 1 , Figure 2 The method for generating chest X-ray reports based on thought chain reasoning and active learning, as shown, includes the following steps:
[0062] S1. Acquire multi-view chest X-ray images and their corresponding reports; perform viewpoint recognition on the multi-view chest X-ray images, generate single-view images, and perform necessary viewpoint classification and preprocessing.
[0063] Data preparation: Select chest X-ray images containing multiple views {I1, I2, ..., I...} N The training dataset consists of samples from the} and its diagnostic report R. Perspectives include, but are not limited to: PA (posteroanterior), AP (anteroposterior), and lateral.
[0064] Viewpoint identification and organization: Images within the same study are classified by viewpoint and standardized into types such as single PA, single AP, single lateral view, AP+lateral view, PA+lateral view, AP+PA, AP+PA+lateral view, etc., to facilitate subsequent fusion and reasoning sequence planning.
[0065] S2. A shared visual encoder is used to extract image features from each viewpoint. Through cross-viewpoint feature fusion, a multi-viewpoint visual token is obtained: V = {v1, v2, ..., v...} M}∈R M×d Where M is the sequence length and d is the feature dimension;
[0066] Multi-view feature encoding: A shared visual encoder Encoder(·) is used to extract sequence features for each viewpoint, and for the i-th viewpoint image I... i :
[0067]
[0068] Among them, f i Let M be the sequence features of the i-th single-view image, and M be the sequence length.
[0069] Cross-view fusion: To obtain compact and fully information-integrated multi-view visual tokens, a centralized multi-head attention mechanism is used for cross-view fusion and residual convergence.
[0070]
[0071] V = {v1, v2, ..., v} M}∈R M×d
[0072] Where N is the number of viewpoints, M is the sequence length, d is the feature dimension, V is the multi-view visual token, and MHA(·) represents the multi-head attention mechanism.
[0073] S3. Construct a systematic hierarchical diagnostic process, and sequentially execute a structured five-stage chain reasoning for each anatomical structure in the chest X-ray, including regional localization, disease identification, severity assessment, probabilistic reasoning, and evidence-based interpretation and summary, to generate structured reasoning prompts.
[0074] A clinically aligned hierarchical diagnostic process: Following the framework of "perspective → anatomical structure," the examination sequence of anatomical structures may include: supporting structures, lung fields, mediastinum, great vessels, bony structures, diaphragm, and at least two or all of other structures. The lung fields can be further subdivided into: left / right lung, right upper / middle / lower lobe, left upper / lower lobe, etc., facilitating regional localization.
[0075] Structured five-stage chain reasoning: For each anatomical structure, a structured intermediate reasoning chain is generated, which sequentially includes:
[0076] (1) Regional positioning;
[0077] (2) Disease identification;
[0078] (3) Severity assessment (e.g., mild, mild-moderate, moderate, moderate-severe, severe, widespread);
[0079] (4) Probabilistic reasoning (probabilistic analysis of comorbidities, potential causes, and related abnormalities);
[0080] (5) Evidence-based explanation and summary (condensed into concise structured statements).
[0081] Cross-structure integrated report generation: Summarize the reasoning summaries of each structure, organize and polish them in a predetermined order, and generate a complete report text.
[0082] S4. Input the multi-perspective visual tokens and structured reasoning hints into the large language model, and generate the target sequence through autoregression.
[0083] Autoregressive Language Modeling and Training Objective: Multi-view visual tokens and structured inference hints are input into a large language model. Target sequences are generated using an autoregressive approach. The training objective is cross-entropy loss. The probability of inference chain generation is modeled using the following formula:
[0084]
[0085] Among them, c t C is the thought chain token output at the current moment. 0:t-1 Let x = concat{S, prompt}, where prompt is the structured reasoning hint for the input large model, T is the length of the thought chain, and L is the training loss function. CACoT The cross-entropy loss function is as follows:
[0086]
[0087] S5. Construct a scoring function based on the joint scoring of uncertainty and inconsistency; in each iteration, label the samples in the unlabeled sample pool according to the sample scores to train the language model for generating chest X-ray reports;
[0088] like Figure 3 The active learning-driven selection of valuable samples and iterative training are illustrated. To reduce the cost of high-quality thought chain annotation and verification, this invention prioritizes high-scoring samples from the unlabeled sample pool based on a joint scoring system of uncertainty and inconsistency, and uses these samples for model training iterations. The scoring function is as follows:
[0089]
[0090] in, For uncertainty, i.e., the entropy average based on the top-1 probability of the generated sequence, the calculation method is as follows:
[0091]
[0092] Where, p t =p(c t |C 0:t-1 x) represents the top-1 probability of generating the t-th token. Inconsistency is indicated by a consistency score based on the degree of overlap of medical entities, calculated as follows:
[0093]
[0094] Among them, E c With E r These are the sets of medical entities extracted from the generated reasoning report and the reference report, respectively.
[0095] In each iteration, the top-scoring samples are selected for mental chain annotation and used for training to update the model, thereby improving training efficiency and performance.
[0096] S6. Input the chest X-ray image to be tested into the trained language model to generate the thought chain and complete report text.
[0097] During the testing phase, multi-view images are input, and multi-view visual tokens are obtained through the above steps. Combined with the prompt of the input large model, a thought chain is generated in an autoregressive manner. The five-stage reasoning results of each anatomical structure are merged in a predetermined order to obtain a complete report text.
[0098] Example 2:
[0099] like Figure 2 As shown, the chest X-ray report generation system based on thought chain reasoning and active learning includes:
[0100] Viewpoint recognition module: performs PA / AP / lateral viewpoint recognition and organization on input images;
[0101] Visual encoding module: Extracts sequential features from images from various viewpoints;
[0102] Cross-view fusion module: It adopts centralized multi-head attention to aggregate multi-view information and outputs multi-view visual tokens;
[0103] Structured Reasoning Module: Constructing Structured Clinical Alignment Reasoning Hints (Five Stages);
[0104] Inference Decoding Module: Based on multi-view visual tokens and structured reasoning hints, it generates reports using an autoregressive approach;
[0105] Active learning selection module: Enables joint scoring of uncertainty and inconsistency and sample selection;
[0106] Training module: Performs cross-entropy loss optimization and active learning iteration;
[0107] Example 3: Data and Annotation Construction
[0108] To obtain a structured and clinically relevant reasoning chain, the present invention can construct training data in the following manner:
[0109] The initial five-stage reasoning text is generated based on a pre-trained large language model, and then reviewed and proofread by senior doctors. Entity extraction tools are used to extract medical entities from the report and reasoning text for consistency measurement and active learning scoring. Through active learning iteration, reasoning annotations are supplemented / improved for samples with high scores, thereby continuously improving model capabilities and controlling annotation costs.
[0110] Example 4: Parameters and Optional Settings
[0111] The severity scale for the five-stage reasoning can be selected as: mild, mild-moderate, moderate, moderate-severe, severe, and extensive; the priority of perspective combinations and the order of anatomical structures can be configured according to specific departmental standards; hyperparameters for active learning, such as α and β, can be set based on validation set performance; the visual encoder and language decoder can adopt mainstream architectures, and the language model can be efficiently fine-tuned through methods such as LoRA.
[0112] Example 5:
[0113] To verify the performance of the CACoT method of this invention, comprehensive quantitative and qualitative experiments were conducted using two publicly available large chest X-ray datasets, MIMIC-CXR and CheXpert Plus. The performance metrics compared included several commonly used medical metrics for report generation and diagnostic accuracy.
[0114] Quantitative performance comparison
[0115] Tables 1 and 2 show the performance comparison of CACoT with eight existing methods on the MIMIC-CXR and CheXpert Plus datasets, respectively. Experimental results show that CACoT significantly outperforms most baseline methods in all major metrics, especially in indicators reflecting the accuracy and clinical rationality of reported diagnoses, such as RaTEScore and LLMScore related to medical entities. On the MIMIC-CXR dataset, CACoT achieves a RaTEScore of 0.526, outperforming recent strong baselines MAIRA-2 (0.525) and LLaVA-Rad (0.521). On the CheXpert Plus dataset, CACoT also achieves the highest RaTEScore (0.506) and the lowest RaCliQ-v1 (1.130), indicating higher authenticity and consistency in its generated reports. Using the active learning strategy designed in this paper, CACoT, trained with only 50% of the labeled training data, still maintains performance close to that of full-scale training, demonstrating excellent data utilization efficiency.
[0116] Ablation analysis
[0117] like Figure 5 Tables 3 and 4 show the ablation experimental results for key components, further revealing the effectiveness of the method design: multi-view feature fusion significantly improved the accuracy and semantic consistency of the report, demonstrating the necessity of cross-view information integration; the introduction of a fine-grained structured diagnostic reasoning chain (five-stage reasoning) and the organization of diagnostic reasoning according to the anatomical structure sequence of clinical practice effectively improved the coherence and medical rationality of the model-generated report; the sample selection strategy jointly guided by uncertainty and inconsistency indicators introduced by active learning effectively maintained and improved model performance, especially with limited labeled data, verifying its advantages in labeling cost control; the high viewpoint recognition accuracy of multi-view images (approximately 98% or more) ensured the orderly unfolding of the diagnostic process.
[0118] Qualitative analysis
[0119] like Figure 4 As shown, compared with reports generated by current mainstream models (MAIRA-2, LLaVA-Rad, MedGemma), the CACoT method of this invention can more comprehensively and accurately identify support devices, lesion locations, severity, and possible causes in chest X-rays. Especially in cases with multiple devices and complex lesions, CACoT generates content covering the vast majority of important abnormalities, reducing false positives and missed diagnoses. The logical order and language structure of the report are more in line with the diagnostic habits of radiologists.
[0120] Overall, CACoT achieves high accuracy and strong interpretability in automated report generation through a clinically aligned hierarchical diagnostic process and structured chain reasoning. Furthermore, its active learning strategy significantly reduces the need for annotation, demonstrating broad application potential.
[0121] Table 1: Performance comparison on MIMIC-CXR; This indicates that active learning was used with only 50% of the training data; the best result is shown in bold; ↑ indicates that the larger the value, the better the performance; ↓ indicates that the smaller the value, the better the performance.
[0122]
[0123] Table 2: Performance comparison on CheXpert Plus; This indicates that active learning was used with only 50% of the training data; the best result is shown in bold; ↑ indicates that the larger the value, the better the performance; ↓ indicates that the smaller the value, the better the performance.
[0124]
[0125]
[0126] Table 3: Ablation study on MIMIC-CXR; perspective, reasoning, uncertainty, and inconsistency represent the application of multi-perspective information fusion, thought chain reasoning, diagnostic order, uncertainty modeling, and inconsistency estimation during training, respectively; best results are indicated in bold; except for the RadCliQ-v1 index, all other indices are considered better the higher the value, while the RadCliQ-v1 index is considered better the lower the value.
[0127]
[0128] Table 4: Ablation studies on CheXpertPlus; perspective, reasoning, uncertainty, and inconsistency represent the application of multi-perspective information fusion, thought chain reasoning, diagnostic order, uncertainty modeling, and inconsistency estimation during training, respectively; best results are indicated in bold; except for the RadCliQ-v1 index, higher values are preferred for all other indices, while lower values are preferred for RadCliQ-v1.
[0129]
Claims
1. A method for generating chest X-ray reports based on thought chain reasoning and active learning, characterized in that, Includes the following steps: S1. Acquire and preprocess multi-view chest X-ray images and their corresponding reports; The multi-view chest X-ray images are subjected to viewpoint identification and viewpoint classification. S2. A shared visual encoder is used to extract sequence features from each viewpoint, and multi-view visual tokens are generated through cross-viewpoint feature fusion. S3. Construct a systematic hierarchical diagnostic process, and sequentially execute structured five-stage chain reasoning for each anatomical structure in multi-view chest X-ray images to generate structured reasoning prompts. S4. Input the multi-view visual token and the structured reasoning prompt into the large language model, and generate the target sequence through autoregression. S5. Construct a scoring function based on the joint scoring of uncertainty and inconsistency; in each iteration, label the samples in the unlabeled sample pool according to the sample scores to train the language model for generating chest X-ray reports; S6. Input the multi-view chest X-ray images to be tested into the trained language model to generate the thought chain and complete report text.
2. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 1, characterized in that, In S2, the sequence features are extracted using an Encoder(·) visual encoder for the i-th single-view image I. i : Among them, f i Let M be the sequence features of the i-th single-view image, and M be the sequence length.
3. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 2, characterized in that, In S2, a centralized multi-head attention mechanism is used for cross-perspective fusion and residual convergence: V={v1,v2,…,v M }∈R M×d Where N is the number of viewpoints, M is the sequence length, d is the feature dimension, V is the multi-view visual token, and MHA(·) represents the multi-head attention mechanism.
4. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 1, characterized in that, To achieve detailed reasoning about each anatomical structure, the structured intermediate reasoning described in S3 includes, in sequence: Region localization: Fine-grained region division and localization of the target structure; Disease identification: Identify potential pathologies or abnormalities and provide high-level interpretations; Severity assessment: The degree of lesion is characterized using a discrete scale; Probabilistic reasoning: Based on contextual clues, provide a probabilistic analysis of comorbidities, interrelated abnormalities, or potential causes; Evidence-based interpretation and summary: Summarize the aforementioned information and output a structured summary statement.
5. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 1, characterized in that, S4 also includes generating inference chain probabilities through an autoregressive approach, modeled as follows: Among them, c t C is the thought chain token output at the current moment. 0:t-1 Let x = concat{V, prompt}, where prompt is the structured reasoning instruction generated from the chest X-ray report of the input large language model, and T is the length of the thought chain.
6. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 5, characterized in that, The training objective of the language model is cross-entropy loss. The cross-entropy loss formula is constructed based on the probability generated by the inference chain as follows:
7. The method for generating chest X-ray reports based on thought chain reasoning and active learning according to claim 1, characterized in that, The scoring function described in S5 is: in, For uncertainty, i.e., the entropy average based on the top-1 probability of the generated sequence, the calculation method is as follows: Where, p t =p(c t |C 0:t-1 x) represents the top-1 probability of generating the t-th token; Inconsistency is indicated by a consistency score based on the degree of overlap of medical entities, calculated as follows: Among them, E c With E r These are the sets of medical entities extracted from the generated reasoning report and the reference report, respectively.
8. A chest X-ray report generation system based on thought chain reasoning and active learning, characterized in that, include: The perspective recognition module is used to acquire multi-view chest X-ray images and their corresponding reports, perform perspective recognition on the multi-view chest X-ray images, and generate single-view images. A visual encoding module is used to extract sequential features from each of the single-view images using a shared visual encoder; A cross-view fusion module is used to fuse the sequence features to generate a multi-view visual token set; The structured reasoning module is used to construct a hierarchical diagnostic process that first determines the perspective and then systematically examines each anatomical structure from that perspective. For each anatomical structure, it performs structured five-stage chain reasoning to generate a structured intermediate reasoning chain. The reasoning and decoding module is used to generate target sequences based on a multi-view visual token set and structured reasoning hints, using an autoregressive approach. The active learning selection module is used to construct a scoring function based on the joint scoring of uncertainty and inconsistency, and to label samples in the unlabeled sample pool according to the sample scores; The training module is used to perform cross-entropy loss optimization and active learning iteration; The chest X-ray report generation system based on thought chain reasoning and active learning is used to implement the chest X-ray report generation method based on thought chain reasoning and active learning as described in claim 1.