Multi-scale and multi-region thyroid nodule prediction method

Abstract

The invention relates to the technical field of medical image processing, and provides a multi-scale multi-region thyroid nodule prediction method, which comprises the following steps: acquiring multi-dimensional image data, synchronously monitoring physiological data, deploying a high-precision clock as a time source, synchronizing equipment to the time source, and outputting external TTL trigger to an ultrasonic host to calibrate the acquisition moment of each frame; carrying out resampling, normalization and registration on the collected data to form a standardized image tensor, inputting the standardized image tensor into a lightweight backbone network, extracting local features of candidate ROI, carrying out cross-regional self-attention and global feature interactive fusion, and forming joint characterization of a focus of focus; inputting the real-time risk probability of each candidate nodule into a joint fusion network, outputting the real-time risk probability of each candidate nodule, performing calibration function processing and weight fusion to obtain an instantaneous probability, and then implementing exponential smoothing to calculate a real-time risk score; and establishing a multi-channel notification mechanism, setting a notification path and a timeout interval, and synchronizing the operation condition to an alarm event to form a judgment closed loop.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically a multi-scale, multi-regional method for predicting thyroid nodules. Background Technology

[0002] Predictive methods for thyroid nodules include combining clinical symptoms, thyroid ultrasound examination, and routine blood tests to assess the risk of benign or malignant nodules. However, for high-risk nodules, ultrasound-guided fine-needle aspiration biopsy is required for diagnosis. Currently, artificial intelligence and big data technologies are being used for prediction, using ultrasound imaging data to help determine the invasiveness and metastatic potential of nodules.

[0003] The Chinese patent application (CN202310215158.X) discloses a multi-scale, multi-regional thyroid nodule prediction method. This method includes: preparing a training dataset; constructing a feature extraction model; constructing a feature reconstruction model; determining the prediction model; and using ultrasound images of different scales and regions as input, combining a high- and low-frequency module with a multi-scale feature extraction module. Specifically, during network training, the feature map is divided into high- and low-frequency features according to channel direction, and features are extracted from thyroid nodules, followed by feature reconstruction to obtain discriminative features. This network provides more comprehensive thyroid nodule feature extraction and achieves higher accuracy in predicting benign or malignant nodules.

[0004] In the field of medical image processing technology, although there are technical solutions that involve preprocessing thyroid ultrasound images provided by hospitals, inputting thyroid nodule images of different scales and regions into a high- and low-frequency feature extraction module, reconstructing features from the multi-scale feature extraction results, and training a network model using a training set, the accuracy of thyroid nodule screening in primary healthcare institutions remains low due to limited resources. This is particularly true for identifying small nodules and early lesions. Furthermore, traditional two-dimensional ultrasound examinations are limited by the operator's experience and lack standardized procedures, leading to significant differences in diagnostic results between different regions and institutions. Therefore, a thyroid nodule prediction method that incorporates multi-scale feature extraction and a region-focused mechanism is needed. Summary of the Invention

[0005] This invention provides a multi-scale, multi-regional method for predicting thyroid nodules, aiming to solve the technical and equipment difficulties faced by primary healthcare institutions in identifying thyroid lesions.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: A multi-scale, multi-regional thyroid nodule prediction method is provided, comprising:

[0007] Collect multidimensional image data, monitor physiological data synchronously, deploy a high-precision clock as a time source, synchronize the ultrasound host, physiological monitor and acquisition workstation to the time source, output external TTL trigger to the ultrasound host to calibrate the acquisition time of each frame, and cache the acquired modal data;

[0008] Multi-scale features are extracted, and multi-region focus is applied to lesions. The collected data is resampled, normalized, and registered to form a standardized image tensor, which is input into a lightweight backbone network. Multi-scale feature maps are constructed based on the feature pyramid network. Local features are extracted from candidate ROIs, and cross-regional self-attention and global feature interaction are performed to form a joint representation of focused lesions.

[0009] By combining algorithmic data fusion, nodule risk is assessed in real time. Multi-scale image features, local attention features, and frame-level physiological scalars are input into the joint fusion network, which outputs the real-time risk probability of each candidate nodule. After processing by a calibration function and fusion according to weights, the instantaneous probability is obtained, and then exponential smoothing is performed to calculate the real-time risk score.

[0010] Establish an intelligent early warning system that automatically issues warnings based on predictions. When the real-time risk score threshold meets predetermined rules, an alarm is triggered. Establish a multi-channel notification mechanism, set notification paths and timeout intervals, and synchronize operational status to form a closed-loop judgment process for alarm events.

[0011] As a preferred implementation method, the specific steps for acquiring multidimensional image data and simultaneously monitoring physiological data are as follows:

[0012] A master clock with a high-precision clock protocol is deployed within the local acquisition network. The ultrasound host is configured as the network slave clock. Simultaneously, the physiological monitoring equipment and the data acquisition workstation are synchronized to the slave clock to form a unified time reference system. The physiological monitoring equipment generates a TTL level signal, which is connected to the external trigger interface of the ultrasound host to provide a hardware acquisition time marker for each frame of ultrasound image. In the image frame acquisition callback, the microsecond-level timer of the acquisition workstation writes a unified timestamp to the current image frame, the ECG waveform sampling points acquired by the physiological monitor, and the attitude samples and contact force samples acquired by the inertial measurement unit at the probe end. The image frame and the timing signals of each modality are written into a circular buffer respectively. Alignment and resampling processing are performed according to the timestamp. The frame event identifier is embedded in the image frame header, and the aligned waveform is encapsulated into an atomic data packet. Each acquisition is evaluated according to the preset sharpness threshold, motion amplitude threshold, and signal integrity threshold. If the evaluation fails, the resampling process is triggered.

[0013] Alignment and resampling are performed based on timestamps. A fixed-length circular buffer is implemented at the acquisition workstation. All image frames, physiological samples, and probe sensor samples are written one by one with microsecond-level timestamps. Neighboring samples are retrieved from the buffer and real-time linear interpolation is performed to obtain physiological scalars and event markers that are precisely corresponding to each frame. The quality score is calculated online, and the weighted average of three components, namely image sharpness, probe motion amplitude, and ECG signal integrity, is calculated comprehensively. Resampling is triggered when the score is below a threshold.

[0014] As a preferred implementation method, the specific steps for extracting multi-scale features and focusing on lesions in multiple regions are as follows:

[0015] Spatial resampling, intensity normalization, and inter-frame registration are performed on the acquired multidimensional images and physiological data. First, the pixel physical size of the images is read, and all images are resampled to the same grid according to the target resolution. The spatial scale is standardized, and the images are aligned with the reference template through histogram matching. Outliers are removed by percentile truncation, and intensity normalization is completed. Based on the attitude data provided by the probe's built-in inertial measurement unit, the initial spatial transformation parameters are calculated, and rigid coarse registration between images is performed. Non-rigid fine registration is performed using a local deformation model to further eliminate spatial deviations caused by probe displacement. The standardized image tensor data is output to a lightweight backbone network. Based on the feature pyramid network, a scale hierarchy is constructed from low to high to obtain multi-scale feature maps. By deploying dilated convolution and ASPP on each scale feature, the receptive field is expanded, and attention mechanisms are introduced in the channel and spatial dimensions. A temporal encoder is used for image frame sequences containing the time dimension to extract dynamic multi-scale temporal features. After applying one-dimensional convolution and gated encoding to the physiological signals, cross-modal attention fusion is performed on multiple scales to obtain a joint feature representation of the fused physiological state and image structure.

[0016] The acquired images are first linearized, and then the gray-level histogram of the image to be corrected within the thyroid mask is mapped to the pre-selected reference template histogram to obtain an image with a gray-level distribution consistent with the reference template. The formula for calculating the histogram is:

[0017] ,

[0018] Where x and y represent the horizontal and vertical coordinates of a pixel, respectively, and r represents the grayscale index. This represents the histogram of the r-th bucket. Indicates pixel intensity value, Represents the mask function;

[0019] To focus on suspected lesion areas and reduce background tissue interference, the entire image was automatically segmented into thyroid glands to obtain gland masks and defined ROIs. Several candidate nodule boxes were generated within the gland mask using a fast candidate detector. Each candidate box was normalized in size and aligned with the region to obtain uniformly sized local patches. A lightweight region attention module was deployed on each local patch. By interacting and fusing each local feature with the corresponding global features of the entire gland under the attention mechanism, joint features were obtained. Lightweight cross-regional self-attention was introduced between candidate regions to realize neighborhood information exchange and weighted summarization of multiple candidate results to obtain lesion representations enhanced by the interrelationships between neighboring regions.

[0020] As a preferred implementation method, the specific steps for real-time assessment of nodule risk by combining algorithmic data fusion are as follows:

[0021] Preprocessed and time-aligned multi-dimensional image multi-scale features, candidate ROI attention-weighted features, and frame-level physiological features are input into a joint fusion network. This network consists of multi-scale convolutional branches, a temporal encoder, and a cross-modal attention fusion layer. A lightweight classification and regression head is configured on the fused features. Model ensemble and probability calibration are used to calculate the real-time risk score and corresponding confidence estimate for each candidate nodule. The real-time risk score calculation formula is as follows:

[0022] ,

[0023] Where t represents time and k represents the candidate nodule. Indicates the real-time risk score. Represents the smoothing coefficient. This represents the instantaneous risk probability of candidate nodule k at time t;

[0024] Real-time risk scores are thresholded and rule-mapped, and time smoothing and short-term consistency are combined for detection. Attention heatmaps, key physiological event labels and uncertainty prompts are output simultaneously in the decision path, and a complete audit log is recorded.

[0025] When the real-time risk score is less than 0.2, the threshold corresponds to low risk, and the rule corresponds to routine follow-up. When the real-time risk score is greater than or equal to 0.2 and less than 0.5, the threshold corresponds to risk requiring observation, and the rule corresponds to ultrasound re-examination at 6 months. When the real-time risk score is greater than 0.5 and less than or equal to 0.8, the threshold corresponds to medium to high risk, and the rule corresponds to recommending fine-needle aspiration for further examination. When the real-time risk score is greater than 0.8, the threshold corresponds to high risk, and the rule corresponds to timely hospitalization.

[0026] As a preferred implementation method, the specific steps for establishing an intelligent early warning system and automatically issuing early warnings based on predictions are as follows:

[0027] Alarm decisions are triggered when the threshold corresponding to the real-time risk score exceeds the required risk level. Medium and high risks are indicated by prompts on the user interface, while high risks are notified to the attending physician via SMS. Alarm messages are encapsulated as structured events and written to the local alarm queue according to priority. The priority is derived by combining the real-time risk score and confidence level with preset weights. By establishing a multi-channel notification mechanism and setting different notification paths and response timeouts according to priority, the results of medical personnel's operations are written back to the alarm event and synchronized to the patient's file, thus obtaining a human-machine collaborative judgment closed loop. The historical sessions and follow-up results of the same patient are automatically compared, and an escalation strategy is triggered for continuously rising nodules. A complete audit log is recorded in the alarm generation and execution path, and alarm performance is statistically evaluated regularly.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This invention improves the ability to identify early micro nodules and reduces the occurrence of missed diagnoses and misdiagnoses by adding physiological data to the imaging data through the fusion analysis of multi-dimensional data.

[0030] 2. This invention obtains geometrically and grayscale-normalized input tensors by performing physical-scale resampling, histogram matching, and pixel-level inter-frame registration on images, thereby reducing the impact of differences in equipment, gain, and acquisition parameters on model performance and improving cross-center generalization. Attached Figure Description

[0031] Figure 1 This is a flowchart of a multi-scale, multi-regional thyroid nodule prediction method.

[0032] Figure 2 This is a comparison chart of the effectiveness of a multi-scale, multi-regional thyroid nodule prediction method. Detailed Implementation

[0033] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0034] Example 1, such as Figure 1 This is a multi-scale, multi-regional method for predicting thyroid nodules, including the following steps:

[0035] Collect multidimensional imaging data and simultaneously monitor physiological data;

[0036] Extract multi-scale features and focus on lesions across multiple regions;

[0037] By combining algorithms and data, the risk of nodules can be assessed in real time.

[0038] Establish an intelligent early warning system to issue automatic warnings based on forecasts.

[0039] The specific implementation steps are as follows: A multi-scale, multi-regional thyroid nodule prediction method, wherein the specific steps for acquiring multi-dimensional imaging data and simultaneously monitoring physiological data are as follows:

[0040] A master clock with a high-precision clock protocol is deployed within the local acquisition network. The ultrasound host is configured as the network slave clock. Simultaneously, the physiological monitoring equipment and the data acquisition workstation are synchronized to the slave clock to form a unified time reference system. The physiological monitoring equipment generates a TTL level signal, which is connected to the external trigger interface of the ultrasound host to provide a hardware acquisition time marker for each frame of ultrasound image. In the image frame acquisition callback, the microsecond-level timer of the acquisition workstation writes a unified timestamp to the current image frame, the ECG waveform sampling points acquired by the physiological monitor, and the attitude samples and contact force samples acquired by the inertial measurement unit at the probe end. The image frame and the timing signals of each modality are written into a circular buffer respectively. Alignment and resampling processing are performed according to the timestamp. The frame event identifier is embedded in the image frame header, and the aligned waveform is encapsulated into an atomic data packet. Each acquisition is evaluated according to the preset sharpness threshold, motion amplitude threshold, and signal integrity threshold. If the evaluation fails, the resampling process is triggered.

[0041] Specifically, the process involves writing image frames and temporal signals of each modality into a circular buffer, performing alignment and resampling based on timestamps, and implementing a fixed-length circular buffer (e.g., 5 seconds) at the acquisition workstation. All image frames, physiological samples, and probe sensor samples are written one by one with microsecond-level timestamps. Neighboring samples are retrieved from the buffer and real-time linear interpolation is performed to obtain physiological scalars and event markers that precisely correspond to each frame. A quality score is calculated online, and the weighted average of three components—image sharpness, probe motion amplitude, and ECG signal integrity—is calculated comprehensively. Resampling is triggered when the score is below a threshold, such as when the sharpness component is below 0.2.

[0042] The specific steps for extracting multi-scale features and focusing on lesions across multiple regions are as follows:

[0043] Spatial resampling, intensity normalization, and inter-frame registration are performed on the acquired multidimensional images and physiological data. First, the pixel physical dimensions of the images are read. All images are resampled to the same grid according to the target resolution, and the spatial scale is standardized. Histogram matching and alignment with a reference template (a dynamic template generated based on a large number of samples) are then performed, with percentile truncation used to remove outliers. Intensity normalization is then completed. Based on the attitude data provided by the probe's built-in inertial measurement unit, initial spatial transformation parameters are calculated for rigid coarse registration between images. Non-rigid fine registration is then performed using a local deformation model to further eliminate spatial deviations caused by probe displacement. Standardized image tensor data is output to a lightweight backbone network. A multi-scale feature map is obtained by constructing a scale hierarchy from low to high based on a feature pyramid network. The feature pyramid network is a convolutional neural network structure that simultaneously generates semantically strong feature maps with varying resolutions at different spatial scales and fuses them. This is achieved by deploying dilated convolutions and dilated spatial pyramid pooling on features at each scale. Pooling (ASPP) expands the receptive field and introduces attention mechanisms in the channel and spatial dimensions. It uses a temporal encoder, such as a preferred temporal convolutional network, to extract dynamic multi-scale temporal features for image frame sequences containing a time dimension. It also performs cross-modal attention fusion on physiological signals after one-dimensional convolution and gating encoding to obtain a joint feature representation of the fused physiological state and image structure.

[0044] Specifically, the process involves linearizing the acquired image through histogram matching and alignment with a reference template, and then mapping the grayscale histogram of the image to be corrected within the thyroid mask to the pre-selected reference template histogram, resulting in an image with a grayscale distribution consistent with the reference template. The formula for calculating the histogram is as follows:

[0045] ,

[0046] Where x and y represent the horizontal and vertical coordinates of a pixel, respectively, and r represents the grayscale index. This represents the histogram of the r-th bucket. Indicates pixel intensity value, Represents the mask function;

[0047] To focus on suspected lesion areas and reduce background tissue interference, the entire image was automatically segmented into thyroid glands to obtain gland masks and define a limited Region of Interest (ROI). Within the gland mask, a fast candidate detector was used to generate several candidate nodule boxes. Each candidate box was normalized in size and aligned with the region to obtain uniformly sized local patches. The fast candidate detector is a lightweight pre-screening module that rapidly extracts features from the entire image, using simple filtering and thresholding to generate highly recalled candidate boxes. A lightweight region attention module, such as channel attention and spatial attention branches, was deployed on each local patch. By interacting and fusing each local feature with the corresponding global features of the entire gland under an attention mechanism, joint features were obtained. Lightweight cross-regional self-attention was introduced between candidate regions to achieve neighborhood information exchange and weighted summarization of multiple candidate results, resulting in a final lesion representation enhanced by the interrelationships between neighboring regions.

[0048] The specific steps for real-time assessment of nodule risk by combining algorithms and data are as follows:

[0049] Preprocessed and time-aligned multi-dimensional image multi-scale features, candidate ROI attention-weighted features, and frame-level physiological features are input into a joint fusion network. This network consists of multi-scale convolutional branches, a temporal encoder, and a cross-modal attention fusion layer. A lightweight classification and regression head is configured on the fused features. Model ensemble and probability calibration are used to calculate the real-time risk score and corresponding confidence estimate for each candidate nodule. The real-time risk score calculation formula is as follows:

[0050] ,

[0051] Where t represents time and k represents the candidate nodule. Indicates the real-time risk score. Represents the smoothing coefficient. This represents the instantaneous risk probability of candidate nodule k at time t;

[0052] Real-time risk scores are thresholded and rule-mapped, and time smoothing and short-term consistency are combined for detection. Attention heatmaps, key physiological event labels and uncertainty prompts are output simultaneously in the decision path, and a complete audit log is recorded.

[0053] Specifically, the real-time risk score is mapped to a threshold and a rule. When the real-time risk score is less than 0.2, the threshold corresponds to low risk and the rule corresponds to routine follow-up. When the real-time risk score is greater than or equal to 0.2 and less than 0.5, the threshold corresponds to risk requiring observation and the rule corresponds to ultrasound re-examination at 6 months. When the real-time risk score is greater than 0.5 and less than or equal to 0.8, the threshold corresponds to medium to high risk and the rule corresponds to recommending fine-needle aspiration for further examination. When the real-time risk score is greater than 0.8, the threshold corresponds to high risk and the rule corresponds to timely hospitalization.

[0054] The specific steps for establishing an intelligent early warning system and issuing automatic warnings based on predictions are as follows:

[0055] The system triggers alarm decisions when the threshold corresponding to the real-time risk score exceeds the required risk level. Medium- and high-risk cases are indicated via the user interface, while high-risk cases are notified to the attending physician via SMS. Alarm messages are encapsulated as structured events, including session ID, nodule coordinates, risk score, confidence level, attention heatmap, and rule-based handling. These are then written to a local alarm queue according to priority, determined by a weighted fusion of the real-time risk score and confidence level. A multi-channel notification mechanism is established, with different notification paths and response timeouts set according to priority. The system writes the results of medical personnel's actions back to the alarm event and synchronizes them to the patient's file, creating a closed-loop human-machine collaborative decision-making process. Historical sessions and follow-up results for the same patient are automatically compared. For continuously rising nodules, escalation strategies are triggered, such as automatic appointment scheduling. A complete audit log is recorded in the alarm generation and execution path, and alarm performance is periodically evaluated, resulting in an intelligent early warning system that is easy for primary healthcare institutions to deploy and operate.

[0056] like Figure 2 A comparative graph showing the performance of a multi-scale, multi-regional thyroid nodule prediction method is presented. The horizontal axis lists key performance indicators, and the vertical axis represents exemplified performance scores ranging from 0 to 100. The aim is to visually demonstrate the expected improvement of the invention in key capabilities compared to typical prior art.

[0057] Example 2, based on Example 1 above, presents a multi-scale, multi-regional thyroid nodule prediction method in a screening scenario at a primary healthcare institution, specifically as follows:

[0058] Step 1: Deploy a high-precision GPS clock module at the acquisition site as the system time source. Align the ultrasound host, physiological monitor, and acquisition workstation with a short handshake. The physiological monitor uses GPS to anchor the timestamp during each sampling and transmits the timestamped sample to the acquisition workstation via a low-latency message queue to obtain a reliable time stream. Inside the acquisition workstation, a priority-based sliding window buffer is used with a buffer duration of 5 seconds. At the receiving end, the microsecond-level offset is estimated using R-peak cross-correlation, and cubic spline resampling is used to align the physiological stream with the image frames to obtain a high-fidelity frame-level physiological scalar. The quality score is calculated in real time, and resampling is triggered when any key component falls below 0.2.

[0059] Step 2: Input the standardized image tensor into the lightweight backbone network, construct multi-layer feature maps from low to high based on the feature pyramid network, deploy dilated convolution and dilated spatial pyramid pooling on multi-scale features to obtain parallel context response, perform thyroid gland segmentation on the whole image, and generate several candidate ROIs within the gland using a lightweight candidate detector, map each candidate ROI to a fixed-size feature block using ROI-Align mapping, apply channel attention and simplified spatial attention branches to each local feature sequentially to obtain local lesion features that are weighted and amplified in the channel and spatial dimensions, introduce a lightweight cross-regional self-attention mechanism between candidate regions to realize neighborhood feature interaction, and perform interactive fusion of the local features of each ROI with the corresponding global features of the whole gland, and use the joint fused features as the input of the discriminant network;

[0060] Step 3: A cross-modal attention fusion module is introduced on multi-scale features and local attention features. Image features are used as queries and physiological coding features are used as key-value inputs to the attention layer. Multi-task output heads are configured on the joint features to achieve joint training of segmentation, detection and probability regression. Multiple lightweight models with different architectures are combined into an ensemble. The temperature scaling parameters of each model are trained on the validation set for probability calibration. The calibration probabilities are weighted and fused according to pre-set weights. The fusion results are exponentially smoothed to obtain the candidate nodule risk score. The risk score is mapped to the clinical decision rule base and combined with the confidence threshold to perform automatic classification and early warning triggering.

[0061] Step four: Record physician operations in the alarm handling interface, and store the operations, corresponding session images, frame-level physiological signals, and subsequent pathological associations in the annotation database. Implement automated quality screening and sample balancing strategies for new samples in the data pipeline, calculate several quantitative performance indicators such as AUC, sensitivity, specificity, and calibration error on the independent validation set, automatically start the training pipeline when the trigger conditions are met, perform regression validation in the isolated test environment, and gradually replace the online model using a canary deployment strategy to obtain the model iteration deployment process. Write the training configuration, hyperparameters, model version, performance indicators, and physician feedback into the audit log and set alarm thresholds for key indicators to ensure traceability of the entire model evolution chain.

[0062] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A multi-scale, multi-regional method for predicting thyroid nodules, characterized in that, include: Collect multidimensional image data, monitor physiological data synchronously, deploy a high-precision clock as a time source, synchronize the ultrasound host, physiological monitor and acquisition workstation to the time source, output external TTL trigger to the ultrasound host to calibrate the acquisition time of each frame, and cache the acquired modal data; Multi-scale features are extracted, and multi-region focus is applied to lesions. The collected data is resampled, normalized, and registered to form a standardized image tensor, which is input into a lightweight backbone network. Multi-scale feature maps are constructed based on the feature pyramid network. Local features are extracted from candidate ROIs, and cross-regional self-attention and global feature interaction are performed to form a joint representation of focused lesions. By combining algorithmic data fusion, nodule risk is assessed in real time. Multi-scale image features, local attention features, and frame-level physiological scalars are input into the joint fusion network, which outputs the real-time risk probability of each candidate nodule. After processing by a calibration function and fusion according to weights, the instantaneous probability is obtained, and then exponential smoothing is performed to calculate the real-time risk score. Establish an intelligent early warning system that automatically issues warnings based on predictions. When the real-time risk score threshold meets predetermined rules, an alarm is triggered. Establish a multi-channel notification mechanism, set notification paths and timeout intervals, and synchronize operational status to form a closed-loop judgment process for alarm events.

2. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 1, characterized in that: The specific steps for collecting multidimensional image data and simultaneously monitoring physiological data are as follows: A master clock with a high-precision clock protocol is deployed within the local acquisition network. The ultrasound host is configured as the network slave clock. Simultaneously, the physiological monitoring equipment and the data acquisition workstation are synchronized to the slave clock to form a unified time reference system. The physiological monitoring equipment generates a TTL level signal, which is connected to the external trigger interface of the ultrasound host to provide a hardware acquisition time marker for each frame of ultrasound image. In the image frame acquisition callback, the microsecond-level timer of the acquisition workstation writes a unified timestamp to the current image frame, the ECG waveform sampling points acquired by the physiological monitor, and the attitude samples and contact force samples acquired by the inertial measurement unit at the probe end. The image frame and the timing signals of each modality are written into a circular buffer respectively. Alignment and resampling processing are performed according to the timestamp. Frame event identifiers are embedded in the image frame header, and the aligned waveform is encapsulated into atomic data packets. Each acquisition is evaluated according to preset sharpness thresholds, motion amplitude thresholds, and signal integrity thresholds. If the evaluation fails, the resampling process is triggered.

3. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 2, characterized in that: The specific steps for acquiring multidimensional image data and simultaneously monitoring physiological data also include: Alignment and resampling are performed based on timestamps. A fixed-length circular buffer is implemented at the acquisition workstation. All image frames, physiological samples, and probe sensor samples are written one by one with microsecond-level timestamps. Neighboring samples are retrieved from the buffer and real-time linear interpolation is performed to obtain physiological scalars and event markers that are precisely corresponding to each frame. The quality score is calculated online, and the weighted average of three components, namely image sharpness, probe motion amplitude, and ECG signal integrity, is calculated comprehensively. Resampling is triggered when the score is below a threshold.

4. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 1, characterized in that: The specific steps for extracting multi-scale features and focusing on lesions across multiple regions are as follows: Spatial resampling, intensity normalization, and inter-frame registration are performed on the acquired multidimensional images and physiological data. First, the pixel physical size of the images is read, and all images are resampled to the same grid according to the target resolution. Spatial scale is standardized, and histogram matching and reference template alignment are used. Outliers are removed by percentile truncation, and intensity normalization is completed. Based on the attitude data provided by the probe's built-in inertial measurement unit, the initial spatial transformation parameters are calculated, and rigid coarse registration between images is performed. Non-rigid fine registration is performed using a local deformation model to further eliminate spatial deviations caused by probe displacement. The standardized image tensor data is output to a lightweight backbone network. Based on the feature pyramid network, a scale hierarchy is constructed from low to high to obtain multi-scale feature maps. By deploying dilated convolution and ASPP on each scale feature, the receptive field is expanded, and attention mechanisms are introduced in the channel and spatial dimensions. A temporal encoder is used for image frame sequences containing the time dimension to extract dynamic multi-scale temporal features. After applying one-dimensional convolution and gated encoding to the physiological signals, cross-modal attention fusion is performed at multiple scales to obtain a joint feature representation of the fused physiological state and image structure.

5. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 4, characterized in that: The specific steps for extracting multi-scale features and focusing on lesions across multiple regions also include: The acquired images are first linearized, and then the gray-level histogram of the image to be corrected within the thyroid mask is mapped to the pre-selected reference template histogram to obtain an image with a gray-level distribution consistent with the reference template. The formula for calculating the histogram is: ， Where x and y represent the horizontal and vertical coordinates of a pixel, respectively, and r represents the grayscale index. This represents the histogram of the r-th bucket. Indicates pixel intensity value, This represents the mask function.

6. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 4, characterized in that: The specific steps for extracting multi-scale features and focusing on lesions across multiple regions also include: To focus on suspected lesion areas and reduce background tissue interference, the entire image was automatically segmented into thyroid glands to obtain gland masks and defined ROIs. Several candidate nodule boxes were generated within the gland mask using a fast candidate detector. Each candidate box was normalized in size and aligned with the region to obtain uniformly sized local patches. A lightweight region attention module was deployed on each local patch. By interacting and fusing each local feature with the corresponding global features of the entire gland under the attention mechanism, joint features were obtained. Lightweight cross-regional self-attention was introduced between candidate regions to realize neighborhood information exchange and weighted summarization of multiple candidate results to obtain lesion representations enhanced by the interrelationships between neighboring regions.

7. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 1, characterized in that: The specific steps for real-time assessment of nodule risk by combining algorithms and data are as follows: Preprocessed and time-aligned multi-dimensional image multi-scale features, candidate ROI attention-weighted features, and frame-level physiological features are input into a joint fusion network. This network consists of multi-scale convolutional branches, a temporal encoder, and a cross-modal attention fusion layer. A lightweight classification and regression head is configured on the fused features. Model ensemble and probability calibration are used to calculate the real-time risk score and corresponding confidence estimate for each candidate nodule. The real-time risk score calculation formula is as follows: ， Where t represents time and k represents the candidate nodule. Indicates the real-time risk score. Represents the smoothing coefficient. This represents the instantaneous risk probability of candidate nodule k at time t.

8. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 7, characterized in that: The specific steps for real-time assessment of nodule risk by combining algorithms and data include: Real-time risk scores are thresholded and rule-mapped, and time smoothing and short-term consistency are combined for detection. Attention heatmaps, key physiological event labels and uncertainty prompts are output simultaneously in the decision path, and a complete audit log is recorded.

9. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 7, characterized in that: The specific steps for real-time assessment of nodule risk by combining algorithms and data include: When the real-time risk score is less than 0.2, the threshold corresponds to low risk, and the rule corresponds to routine follow-up. When the real-time risk score is greater than or equal to 0.2 and less than 0.5, the threshold corresponds to risk requiring observation, and the rule corresponds to ultrasound re-examination at 6 months. When the real-time risk score is greater than 0.5 and less than or equal to 0.8, the threshold corresponds to medium to high risk, and the rule corresponds to recommending fine-needle aspiration for further examination. When the real-time risk score is greater than 0.8, the threshold corresponds to high risk, and the rule corresponds to timely hospitalization.

10. The method for predicting thyroid nodules at multiple scales and in multiple regions according to claim 1, characterized in that: The specific steps for establishing an intelligent early warning system and issuing automatic warnings based on predictions are as follows: When the threshold corresponding to the real-time risk score exceeds the risk level requiring observation, an alarm decision is triggered. For medium- and high-risk cases, a prompt is displayed on the user interface, while for high-risk cases, a notification is sent to the attending physician via SMS. Alarm messages are encapsulated as structured events and written to the local alarm queue according to priority. The priority is derived by fusing the real-time risk score and confidence level with preset weights. By establishing a multi-channel notification mechanism and setting different notification paths and response timeouts according to priority, the results of medical personnel's operations are written back to the alarm event and synchronized to the patient's file, thus obtaining a closed loop of human-machine collaborative judgment. The historical sessions and follow-up results of the same patient are automatically compared, and an escalation strategy is triggered for continuously rising nodules. A complete audit log is recorded in the alarm generation and execution path, and alarm performance is statistically evaluated regularly.

Images