Contrastive semi-supervised learning based lung nodule dynamic screening method

Abstract

The application provides a lung nodule dynamic screening method based on comparative semi-supervised learning, and belongs to the technical field of medical intelligent diagnosis. The lung nodule dynamic screening method based on comparative semi-supervised learning comprises the following steps: step one: collecting multiple follow-up lung CT images of the same patient, pre-processing the lung region of all CT images, uniformly mapping multiple CTs to a baseline CT space coordinate system by using a three-dimensional registration algorithm, dividing a small number of labeled samples and a large number of unlabeled samples, constructing a time sequence comparative sample pair, and obtaining a comparative semi-supervised data set. The application can complete model training only by using a small number of labeled samples, greatly reduces the labeling cost, fully utilizes the inventory of unlabeled follow-up CT in the hospital, can automatically compare multiple images to identify progressive malignant nodules, significantly reduces the missed detection rate of small nodules and false positives, is suitable for the standardization process of clinical lung cancer early screening and long-term follow-up, and has good clinical popularization and industrialization value.

Description

Technical Field

[0001] This invention belongs to the field of medical intelligent diagnostic technology, specifically a dynamic screening method for lung nodules based on comparative semi-supervised learning. Background Technology

[0002] Lung cancer is the malignant tumor with the highest incidence and mortality rate in my country. Early screening and dynamic follow-up of pulmonary nodules are the core means to reduce the mortality rate of lung cancer. Clinical reliance on radiologists to manually interpret images has problems such as low efficiency, easy to miss small ground-glass nodules, time-consuming comparison of multi-phase images, and large differences in subjective interpretation.

[0003] Existing AI-based lung nodule screening models mostly employ purely supervised learning, heavily relying on manual annotation of CT samples. However, accurate annotation of lung CT scans requires time-consuming work by experienced physicians, resulting in extremely high annotation costs. Furthermore, traditional models only detect single static slices and cannot automatically compare patients' follow-up images. This makes it difficult to distinguish between transient inflammatory nodules and malignant progressive nodules that continue to enlarge and increase in density. In addition, they are prone to misidentifying blood vessels and bronchial cross-sections as nodules, leading to excessively high false positive rates and significantly increasing the workload of doctors in reviewing these images.

[0004] Therefore, this invention proposes a dynamic screening method for pulmonary nodules based on contrastive semi-supervised learning to compensate for and improve the shortcomings of existing technologies. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a dynamic screening method for lung nodules based on contrastive semi-supervised learning.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A dynamic screening method for lung nodules based on contrastive semi-supervised learning includes the following steps: Step 1: Collect lung CT images from multiple follow-up phases of the same patient, perform lung region preprocessing on all CT images, and use a three-dimensional registration algorithm to uniformly map the multiple phase CT images to the baseline CT spatial coordinate system. Divide the images into a small number of labeled samples and a large number of unlabeled samples, construct time-series comparison sample pairs, and obtain a comparison semi-supervised dataset. Step 2: Construct a dual-branch contrastive semi-supervised feature extraction network, including an labeled supervised branch and an unlabeled temporal contrastive branch; the labeled branch completes supervised learning of nodule features based on labeled samples; the unlabeled branch introduces temporal contrastive loss to constrain the differences in lung tissue features at different times of the same patient, generates pseudo-labels through momentum updates and filters low-confidence noise pseudo-labels, and outputs the full-image candidate regions for suspected nodules. Step 3: Based on the registered unified spatial coordinates, the same nodule is matched across time series, and the dynamic changes in nodule volume, CT density, and malignant signs are automatically quantified. The dynamic classification results of low-risk / intermediate-risk / high-risk are output through the risk scoring model. Step 4: Based on the prior anatomical mask of pulmonary vessels, bronchi, and pleura, noise reduction is performed on candidate lesions to eliminate false positive lesions caused by blood vessels and bronchi, and to obtain the true pulmonary nodule screening results. Step 5: Generate a standardized dynamic follow-up screening report based on the nodule location, size, temporal dynamic changes, and risk classification.

[0007] Preferably, in step one, the three-dimensional registration algorithm uses a combination of rigid transformation and elastic deformation to correct spatial misalignment caused by the patient's breathing and body position shift; the time-series comparison sample pairs are composed of the same patient's baseline CT and each follow-up CT paired together.

[0008] Preferably, in step two, the temporal comparison loss constraint rule is as follows: the feature vector distance between normal lung tissue, blood vessels, and bronchi at different times of the same patient is reduced; the feature vector distance between the two images is increased in areas with nodule growth and increased density.

[0009] Preferably, in step two, the pseudo-label filtering method is as follows: set a fixed confidence threshold, retain only candidate nodule pseudo-labels with a confidence level greater than the threshold to participate in network iterative training, and filter out low-confidence noise pseudo-labels.

[0010] Preferably, in step three, the cross-temporal nodule matching determination condition is: if the distance between the baseline and the spatial center of the follow-up candidate nodule is less than a preset threshold, and the cosine similarity of the nodule's three-dimensional features is greater than the similarity threshold, then it is determined to be the same nodule and a unique temporal ID is assigned.

[0011] Preferably, in step three, the quantified indicators of dynamic changes in nodules include: nodule volume change rate, change in average CT value, change in the proportion of solid components, lobulation sign, spiculation sign, and the addition or disappearance of pleural traction signs.

[0012] Preferably, in step four, the anatomical prior noise reduction specifically includes: removing tubular candidate lesions that overlap extensively with blood vessel and bronchial masks, removing pleural patch artifacts without three-dimensional volume, and removing tiny noise candidate regions with a diameter smaller than a preset threshold.

[0013] Preferably, the method is applicable to low-dose lung CT and multi-slice spiral enhanced CT, and is used for early lung cancer screening, multi-cycle follow-up examinations of patients, and AI-assisted diagnosis of batch physical examination images.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: Compared to traditional static supervised lung nodule detection algorithms, this invention has four outstanding technical advantages: 1. Significantly reduce the dependence on annotation and make full use of unlabeled follow-up data. Using a time-series comparative semi-supervised learning framework, only a small number of manually labeled samples are needed to complete model training, reducing data annotation costs by more than 60%. The massive amount of unlabeled follow-up CT data in hospitals can all be used for training, significantly improving data utilization. The model has a stronger generalization ability on CT with different equipment and different slice thicknesses, and the rate of missed detection of small nodules is reduced by 25%.

[0015] 2. Supports multi-time-series dynamic comparison and accurately identifies progressive malignant nodules. Through automatic registration of multi-phase CT and cross-time-series nodule matching, it quantifies the dynamic changes in nodule volume, density, and malignant signs. It can automatically distinguish between short-term shrinking inflammatory nodules and continuously progressing high-risk nodules, making up for the shortcomings of traditional single-phase AI in dynamic evaluation and assisting doctors to identify high-risk lung cancer lesions earlier.

[0016] 3. Anatomical prior noise reduction significantly reduces false positives. Combined with anatomical masking of blood vessels, bronchi, and pleura, it filters out false positive lesions, reduces misdiagnosis caused by blood vessels and trachea, greatly reduces the workload of physicians' manual review, and improves the efficiency of clinical image reading.

[0017] 4. Fully automated process, adaptable to standardized clinical follow-up processes. From multi-phase CT input, registration, nodule detection, time-series quantification to risk grading reports, the entire process is fully automated without manual intervention. It is suitable for various clinical scenarios such as batch screening in physical examinations, long-term follow-up of outpatients, and AI-assisted diagnosis in primary hospitals, and has high industrialization value. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of the present invention; Detailed Implementation

[0020] To make the technical means, creative features, achieved objectives, and effects of this invention readily understandable, the invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope defined by this application.

[0021] Dataset preprocessing and time series registration We selected data from 1000 patients with low-dose lung CT scans, including 2-4 follow-up records. The CT slice thickness was 1 mm, and the resolution was 512×512. Only 100 cases with fine nodule annotations completed by senior radiologists were selected as the annotated sample set; the remaining 900 cases without any manual annotations were selected as the unannotated follow-up sample.

[0022] All CT scans underwent uniform preprocessing: threshold segmentation was used to extract lung parenchyma masks, and irrelevant areas of the thoracic skeleton, muscles, and mediastinum were cropped. A three-dimensional elastic registration algorithm was used to align all follow-up CT scans at 3 months, 6 months, and 12 months to the coordinate system of the initial baseline CT scan, correcting for spatial deviations caused by respiratory displacement and body rotation. After registration, the pixel coordinates of the same anatomical structure were unified. The baseline CT scan of the same patient was paired with each follow-up CT scan to construct time-series comparison sample pairs.

[0023] Two-branch contrastive semi-supervised network training The network backbone uses 3D ResNet as the three-dimensional feature encoder and consists of two parallel branches: 1. The labeled supervised branch takes labeled CT samples as input, and the loss function includes nodule classification loss and bounding box regression loss to learn the basic features of nodules; 2. Input time-paired CT without labeling contrast branches and construct time-contrast loss: calculate the Euclidean distance for feature vectors of the same spatial location in two images, minimize the feature distance of normal tissue, and maximize the feature distance of nodule regions that have enlarged / increased in density; 3. A momentum teacher model is used to generate pseudo-labels for candidate nodules in unlabeled samples. A confidence threshold of 0.7 is set, and only pseudo-labels with a confidence of ≥0.7 are retained for training, while low-confidence noise pseudo-labels are filtered out. After 80 rounds of iterative training, the model converges and outputs 3D candidate bounding boxes for suspected nodules in the full image.

[0024] Cross-time series node matching and dynamic risk quantification Based on the unified spatial coordinates after registration, candidate nodules from baseline CT and follow-up CT outputs are matched: those with a spatial center distance of less than 3 mm and a feature cosine similarity greater than 0.8 are determined to be the same nodule and assigned a unique temporal ID.

[0025] Automatically calculate dynamic indicators: nodule diameter change rate, volume growth rate, average CT value difference, change in the proportion of solid component, and whether new spiculation / lobulation signs have appeared.

[0026] Input dynamic change indicators into the trained risk scoring model: volume increase >20%, increase in solid components, and new spiculation signs are judged as high risk; small volume fluctuations and no malignant signs are judged as medium risk; volume reduction and density decrease are judged as low risk.

[0027] Dissection prior to post-processing noise reduction Using the anatomical masks of blood vessels, bronchi, and pleura extracted during the preprocessing stage, all candidate nodules are filtered: 1. If more than 80% of the candidate box overlaps with the blood vessel / bronchus mask, it will be directly rejected; 2. Remove sheet-like areas that are only attached to the pleura and have no three-dimensional volume; 3. Candidate boxes with tiny noise (less than 2mm in diameter) were removed. The number of false positives decreased by an average of 65% after noise reduction.

[0028] Output screening report and effectiveness verification It automatically summarizes the size, density, dynamic changes, and risk classification of nodules in each follow-up period and generates a standardized AI screening report.

[0029] Clinical test results: Compared with the traditional pure supervised single-phase detection model, the false negative rate of small nodules in this embodiment is reduced by 26%, and the number of false positives is reduced by 63%; the detection accuracy of traditional algorithms using all labeled samples can be achieved using only 10% of the labeled samples; it can automatically identify progressive malignant nodules, meeting the clinical requirements for early screening of lung cancer follow-up.

[0030] In addition to low-dose CT of the lungs, this method can be adapted to multi-slice spiral enhanced CT, and universal screening can be achieved by only fine-tuning the registration parameters and risk score thresholds.

[0031] In the description of this invention, it should be understood that the terms "upper," "side," "inner," etc., indicating orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the elements referred to must have a specific orientation, or be constructed and operated in a specific orientation. In addition, it should be noted that unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal communication of two elements or the interaction relationship between two elements. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic screening method for pulmonary nodules based on contrastive semi-supervised learning, characterized in that, Includes the following steps: Step 1: Collect lung CT images from multiple follow-up phases of the same patient, perform lung region preprocessing on all CT images, and use a three-dimensional registration algorithm to uniformly map the multiple phase CT images to the baseline CT spatial coordinate system. Divide the images into a small number of labeled samples and a large number of unlabeled samples, construct time-series comparison sample pairs, and obtain a comparison semi-supervised dataset. Step 2: Construct a dual-branch comparative semi-supervised feature extraction network, including a labeled supervised branch and an unlabeled temporal comparison branch; the labeled branch completes supervised learning of nodule features based on labeled samples; The unlabeled branch introduces temporal contrast loss to constrain the differences in lung tissue characteristics of the same patient at different times. Pseudo-labels are generated through momentum update and low-confidence noise pseudo-labels are filtered out to output the candidate regions of suspected nodules in the whole image. Step 3: Based on the registered unified spatial coordinates, the same nodule is matched across time series, and the dynamic changes in nodule volume, CT density, and malignant signs are automatically quantified. The dynamic classification results of low-risk / intermediate-risk / high-risk are output through the risk scoring model. Step 4: Based on the prior anatomical mask of pulmonary vessels, bronchi, and pleura, noise reduction is performed on candidate lesions to eliminate false positive lesions caused by blood vessels and bronchi, and to obtain the true pulmonary nodule screening results. Step 5: Generate a standardized dynamic follow-up screening report based on the nodule location, size, temporal dynamic changes, and risk classification.

2. The method for dynamic screening of lung nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step one, the three-dimensional registration algorithm uses a combination of rigid transformation and elastic deformation to correct spatial misalignment caused by the patient's breathing and body position shift; the time-series comparison sample pairs are composed of the same patient's baseline CT and each follow-up CT.

3. The method for dynamic screening of lung nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step two, the temporal comparison loss constraint rule is as follows: the feature vector distance between normal lung tissue, blood vessels, and bronchi at different times of the same patient is reduced; the feature vector distance between the two images is increased in areas with nodule growth and increased density.

4. The method for dynamic screening of lung nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step two, the pseudo-label filtering method is as follows: a fixed confidence threshold is set, and only candidate nodule pseudo-labels with a confidence level greater than the threshold are retained to participate in network iterative training, while low-confidence noise pseudo-labels are filtered out.

5. The method for dynamic screening of lung nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step three, the cross-temporal nodule matching determination condition is: if the distance between the baseline and the spatial center of the follow-up candidate nodule is less than a preset threshold, and the cosine similarity of the nodule's three-dimensional features is greater than the similarity threshold, then it is determined to be the same nodule and a unique temporal ID is assigned.

6. The method for dynamic screening of pulmonary nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step three, the quantitative indicators of dynamic changes in nodules include: nodule volume change rate, average CT value change, change in the proportion of solid components, lobulation sign, spiculation sign, and the addition or disappearance of pleural traction signs.

7. The method for dynamic screening of lung nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, In step four, the anatomical prior noise reduction specifically includes: removing tubular candidate lesions that overlap extensively with blood vessel and bronchial masks, removing pleural patch artifacts without three-dimensional volume, and removing tiny noise candidate regions with a diameter smaller than a preset threshold.

8. The method for dynamic screening of pulmonary nodules based on contrastive semi-supervised learning according to claim 1, characterized in that, The method is applicable to low-dose lung CT and multi-slice spiral enhanced CT, and can be used for early lung cancer screening, multi-cycle follow-up examinations of patients, and AI-assisted diagnosis of batch physical examination images.

Images