Intelligent auxiliary diagnosis method and system for frozen section in thyroid cancer operation

Abstract

The invention relates to the technical field of medical image and artificial intelligence technology crossing, in particular to an intelligent auxiliary diagnosis method and system for a frozen section in a thyroid cancer operation. According to the technical scheme, the method comprises the following steps that a thyroid tissue intraoperative frozen section to be diagnosed is prepared, and the frozen section does not need to be dyed or subjected to section sealing treatment; and scanning the frozen section by using a terahertz near-field imaging system to obtain terahertz near-field image data of the frozen section. According to the method, the terahertz near-field imaging technology is adopted, the situation that cell nucleus morphology is difficult to observe due to ice crystal artifacts in a traditional frozen section is avoided, cancer cells are automatically recognized in combination with an AI model trained based on large-scale labeling data, and the objectivity and consistency of diagnosis are remarkably improved; meanwhile, the steps of dyeing and mounting are omitted, and rapid imaging and analysis are realized, so that the intraoperative diagnosis time is greatly shortened, and a generated visual contour report also provides visual decision support for pathologists.

Description

Technical Field

[0001] This invention relates to the field of medical imaging and artificial intelligence technology, and in particular to an intelligent assisted diagnosis method and system for intraoperative frozen section in thyroid cancer. Background Technology

[0002] Thyroid cancer is the most common malignant tumor of the endocrine system, with papillary thyroid carcinoma accounting for more than 85% of all cases. Intraoperative frozen section pathological examination is a crucial step in thyroid cancer surgery, as its results directly determine the extent of the surgery and subsequent treatment plan.

[0003] Currently, the most widely used intraoperative frozen section diagnostic technique in clinical practice is based on optical microscopy observation of frozen sections stained with hematoxylin and eosin (H&E). However, this technique has the following significant limitations:

[0004] (1) Interference from ice crystal artifacts: During rapid freezing, ice crystals easily form in tissues, causing cell nuclei to swell and become blurred, making it difficult to clearly identify key diagnostic features (such as ground-glass nuclei, nuclear grooves, and intranuclear pseudo-inclusion bodies). Studies have shown that the diagnostic concordance rate between frozen sections and final paraffin sections of papillary thyroid carcinoma is about 85%, with most diagnostic errors stemming from the influence of ice crystal artifacts.

[0005] (2) High degree of subjectivity in diagnosis: The diagnosis results are highly dependent on the experience and subjective judgment of the pathologist. Different doctors or even the same doctor may give inconsistent diagnostic opinions at different times, which affects the consistency and reliability of the diagnosis.

[0006] (3) High time pressure: The intraoperative frozen section needs to be completed within 10-20 minutes for slide preparation and diagnosis. The surgical team usually waits for the pathology report within 30 minutes. The tight time can easily lead to diagnostic omissions or misjudgments.

[0007] In recent years, artificial intelligence technology has been gradually applied to digital pathological image analysis, but its mainstream methods are still based on traditional stained optical images. These images are inherently limited by multiple factors such as staining uniformity, section thickness, and slide preparation quality, making it difficult to overcome the inherent morphological observation limitations of frozen sections.

[0008] Terahertz waves (frequency range 100 GHz–10 THz) are highly sensitive to the vibrational modes of water, polar molecules, and biomacromolecules in biological tissues, providing biochemical and physical characteristic information distinct from optical images. Terahertz near-field imaging technology can overcome the optical diffraction limit, achieving nanometer-level spatial resolution and possessing the potential to characterize cellular and subcellular structures. However, currently, there is no systematic solution combining terahertz near-field imaging technology with artificial intelligence for rapid and objective auxiliary diagnosis of intraoperative frozen sections in thyroid cancer. Therefore, this application proposes an intelligent auxiliary diagnostic method and system for intraoperative frozen sections in thyroid cancer. Summary of the Invention

[0009] The purpose of this invention is to address the current lack of a systematic solution in the art that combines terahertz near-field imaging technology with artificial intelligence for rapid and objective auxiliary diagnosis of intraoperative frozen sections in thyroid cancer, and to propose an intelligent auxiliary diagnostic method and system for intraoperative frozen sections in thyroid cancer.

[0010] In a first aspect, the present invention provides an intelligent assisted diagnosis method for intraoperative frozen section analysis of thyroid cancer, comprising the following steps:

[0011] S1. Prepare intraoperative frozen sections of thyroid tissue to be diagnosed, wherein the frozen sections are not stained or mounted.

[0012] S2. Scan the frozen section using a terahertz near-field imaging system to obtain terahertz near-field image data of the frozen section;

[0013] S3. Input the terahertz near-field image data into a pre-trained thyroid cancer cell recognition model, and have the model output a segmentation result map corresponding to the cancer cell region in the frozen section; wherein, the thyroid cancer cell recognition model is obtained by training a deep learning network based on a training dataset containing a large number of terahertz near-field images of thyroid tissue and corresponding pixel-level cancer cell contour annotations.

[0014] S4. Based on the segmentation result image, generate a visualization report to assist in intraoperative pathological diagnosis.

[0015] Optionally, the terahertz near-field imaging system includes a terahertz source, a focusing lens group for focusing terahertz waves, a metal probe for near-field detection, a detector, and a signal processing and imaging module.

[0016] Optionally, the pixel-level cancer cell contour annotations in the dataset are generated based on the diagnostic results of the corresponding paraffin sections.

[0017] Optionally, the deep learning network is a convolutional neural network with an encoder-decoder structure, and the training dataset contains terahertz near-field images and their annotations of no less than 1,000 cases of thyroid cancer.

[0018] Optionally, the convolutional neural network with encoder-decoder structure is U-Net or a variant thereof.

[0019] In a second aspect, the present invention provides an intelligent assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer for implementing the method described in the first aspect, comprising:

[0020] The terahertz near-field imaging module is used to scan frozen sections of thyroid tissue that have not been stained and mounted, and to acquire terahertz near-field image data at nanometer resolution.

[0021] The data processing and analysis module is communicatively connected to the terahertz near-field imaging module and integrates a pre-trained thyroid cancer cell recognition model. The data processing and analysis module is used to receive the terahertz near-field image data, use the thyroid cancer cell recognition model to automatically identify and segment cancer cell regions in the image, and generate a visual diagnostic auxiliary report based on the segmentation results.

[0022] Optionally, the thyroid cancer cell recognition model is constructed in the following manner:

[0023] A dataset consisting of multiple terahertz near-field images of thyroid tissue and their pixel-level cancer cell contour annotations was obtained;

[0024] The dataset is used to train an initial deep learning network, which is used to perform image semantic segmentation tasks.

[0025] When the training reaches the convergence condition, the thyroid cancer cell recognition model is obtained.

[0026] Optionally, the construction of the dataset includes: using an annotation tool to outline the cancer cell region on the terahertz near-field image based on the gold standard diagnostic results of the corresponding paraffin sections, and generating an annotation mask that corresponds one-to-one with the pixels of the original image.

[0027] Optionally, the terahertz near-field imaging module is based on terahertz scattering scanning near-field optical microscopy technology with a metal probe tip.

[0028] Optionally, the system also includes a display module for presenting the visualized diagnostic aid report to a pathologist, the report including results of overlaying the contours of the cancer region onto the original terahertz near-field image.

[0029] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0030] Terahertz near-field imaging technology is used to directly acquire images of unstained frozen sections, avoiding interference with cell nucleus morphology caused by ice crystals generated during rapid freezing, thereby improving the identification of key features of thyroid cancer cell nuclei.

[0031] Automatic identification and segmentation are achieved by using deep learning models trained on large-scale labeled data, reducing reliance on pathologists' subjective experience and improving the consistency and reproducibility of diagnoses.

[0032] Eliminating the traditional steps of frozen section staining and mounting, and combining rapid terahertz scanning with AI real-time analysis, the entire process can be completed in minutes, meeting the time requirements for rapid intraoperative pathological diagnosis.

[0033] The system automatically generates a visual report containing an overlay image of the cancerous area outline, helping pathologists quickly locate suspicious areas and improve diagnostic efficiency and accuracy.

[0034] This system can provide standardized auxiliary diagnostic support for primary hospitals or pathologists with relatively little experience, and helps to improve the overall level of intraoperative frozen section diagnosis for thyroid cancer.

[0035] In summary, this invention, by employing terahertz near-field imaging technology, avoids the difficulty in observing cell nucleus morphology caused by ice crystal artifacts in traditional frozen sections. Combined with an AI model trained on large-scale labeled data, it achieves automatic identification of cancer cells, significantly improving the objectivity and consistency of diagnosis. At the same time, by omitting staining and mounting steps and achieving rapid imaging and analysis, it greatly shortens the intraoperative diagnostic time, and the generated visual contour report also provides pathologists with intuitive decision support. Attached Figure Description

[0036] Figure 1 Terahertz near-field imaging surface image of a frozen section of papillary thyroid carcinoma;

[0037] Figure 2 Internal image of a frozen section of papillary thyroid carcinoma using terahertz near-field imaging;

[0038] Figure 3 The output image shows the automatically identified cancerous areas. Detailed Implementation

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0040] Example: In a first aspect, the present invention proposes an intelligent assisted diagnostic system for intraoperative frozen section of thyroid cancer, which includes a terahertz near-field imaging module, a data preprocessing and annotation module, and an artificial intelligence analysis and recognition module. The modules are described in detail below.

[0041] 1. Terahertz Near-Field Imaging Module: Terahertz waves have unique advantages in imaging, as they can penetrate non-conductive materials without damage and achieve nanoscale high resolution. Near-field imaging, by probing extremely close to the sample surface, has the core advantage of breaking the diffraction limit, thereby achieving higher resolution. The terahertz near-field imaging system used in this invention is based on THz s-SNOM technology at the tip of a metal probe, possessing nanoscale spatial resolution.

[0042] The terahertz near-field imaging system of the present invention consists of a terahertz source, a mirror assembly, a probe, and other devices. The system's workflow is as follows: the radiation generated by the terahertz source is transmitted and focused by the mirror assembly, and then concentrated at the tip of the probe to generate a near-field interaction with the surface of the frozen slice sample; this interaction modulates the terahertz wave, and the modulated reflected wave returns along the original path of the tip and the mirror assembly and is captured by the detector; finally, the system processes the signal to generate surface and internal images of the sample.

[0043] The testing procedure of this invention is as follows: A frozen section of thyroid tissue to be tested is placed on the stage and fixed in place. The laser is turned on and adjusted to the optimal position. Frequency modulation is performed, and then the magnification of the atomic force microscope is adjusted. The section is moved until the area to be tested is located. The scanning range and frequency are set, and then the needle is inserted to begin scanning. After the scan is completed, surface and internal images of the tested area are obtained. Figure 1 and Figure 2 The images shown are the surface and internal images of a frozen section of papillary thyroid carcinoma obtained by terahertz near-field imaging.

[0044] 2. Data Preprocessing and Labeling Module: Preprocesses the acquired terahertz raw data, including noise reduction, calibration, and image registration.

[0045] Building a high-quality training database: We have collaborated with partner hospitals to collect and annotate thousands of terahertz frozen section images of thyroid cancer and its corresponding benign tissues. Based on the gold standard paraffin section diagnostic results, relevant technicians precisely outlined the thyroid cancer cell regions on the corresponding frozen section terahertz images, thus creating reliable training labels.

[0046] 3. Artificial Intelligence Analysis and Recognition Module: This is the core of the invention for achieving automatic diagnosis. It is built on a large-scale, high-precision labeled dataset and uses a deep convolutional neural network model to deeply integrate the features learned from terahertz images with the diagnostic knowledge of pathologists.

[0047] 3.1 Dataset Construction and Preprocessing

[0048] The model training of this invention relies on a large-scale, pixel-level labeled terahertz image thyroid cancer dataset, the construction process of which is as follows:

[0049] High-precision contour annotation: In collaboration with senior pathologists, we used the Labelme open-source annotation tool to perform pixel-level semantic segmentation annotation on thousands of terahertz near-field images of thyroid tissue. Testers acquired terahertz near-field images of thousands of unstained, unmounted frozen thyroid cancer slides. Pathologists then precisely delineated the contours of each cancerous region on each terahertz near-field image based on the diagnostic results of the corresponding paraffin sections. This annotation method generated annotation masks that corresponded one-to-one with the pixels of the original image, where the background and cancerous regions had different pixel values ​​(e.g., 0 and 255).

[0050] Data format standardization: The JSON files generated by the annotations are batch converted into binary mask images required for model training, forming a paired dataset of "raw terahertz images - annotated masks".

[0051] Data augmentation and preprocessing: To improve the robustness and generalization ability of the model, online and offline data augmentation operations were performed on the training set data to simulate different imaging conditions and tissue differences.

[0052] 3.2 Model Architecture and Training

[0053] Core Network Selection: Given that the labeled data contains fine contour information, this invention employs a deep learning model architecture suitable for image segmentation tasks, using a convolutional neural network based on U-Net or its variants as the core backbone. The encoder-decoder structure and skip connections of this architecture effectively combine deep semantic features with shallow contour information of the image, thereby achieving precise pixel-level segmentation of thyroid cancer cell regions.

[0054] Model input and output: The model input is a preprocessed single-channel or multi-channel terahertz image; the model output is a segmentation prediction map of the same size as the input image, where the value of each pixel represents the probability that the location belongs to a cancer cell (probability value ranges from 0 to 1).

[0055] Loss function and training strategy: A composite loss function combining Dice Loss and Binary Cross-Entropy Loss is employed to force the model to focus on learning the boundaries of the cancer region. End-to-end supervised training is performed using either the Adam or SGD optimizer on a dataset consisting of a large amount of labeled data until the model converges.

[0056] 3.3 Model Functions

[0057] Automatic segmentation and recognition: The trained model can receive a new, undiagnosed thyroid frozen section terahertz near-field image and automatically output its segmentation results. This invention's model can not only identify the presence of cancer cells in the image but also accurately delineate the specific contour, location, and extent of each cancer lesion. For example, [the model can...] Figure 1 and Figure 2 When input into the model, the model automatically identifies and labels cancerous areas in the image, such as... Figure 3 .

[0058] Analysis report generation: The system can automatically analyze the frozen section of the thyroid gland to determine whether it is cancerous, and overlay the outline of the cancerous area onto the terahertz near-field image, providing pathologists with intuitive visual assistance and precisely guiding them to focus on key areas.

[0059] Secondly, the present invention provides a diagnostic method based on the above system, comprising the following steps:

[0060] Step 1: Obtain a thyroid tissue sample to be diagnosed during surgery and prepare frozen sections. It is worth noting that conventional frozen sections are prepared through a process including sampling, rapid embedding, freezing, fine sectioning, brief fixation, staining, and finally mounting. However, the frozen sections used in this invention do not require the staining and mounting steps.

[0061] Step 2: Use a terahertz near-field imaging system to quickly scan the unstained and unmounted frozen thyroid tissue section to obtain terahertz near-field image data of the surface and interior of the frozen section.

[0062] Step 3: Input the terahertz near-field image data obtained by scanning into the trained AI recognition model.

[0063] Step 4: The AI ​​recognition model will automatically identify and segment cancer cell regions in the image and generate a corresponding diagnostic auxiliary report.

[0064] Step 5: The pathologist, combining the aforementioned auxiliary reports with their own experience, makes a final diagnosis and reports it to the operating room.

[0065] It is worth noting that this invention creatively applies terahertz near-field imaging technology to the rapid diagnosis of frozen sections during thyroid cancer surgery. Traditional optical microscopy imaging is severely affected by ice crystal artifacts, while terahertz waves are sensitive to the vibrations of water molecules and biomolecules in biological tissues, enabling direct imaging of unstained and unmounted frozen sections. This physically avoids the interference of ice crystal artifacts on the observation of cell nucleus morphology, thereby obtaining clearer information on cell and subcellular structures.

[0066] This invention eliminates the staining and mounting steps required for traditional intraoperative frozen sections, shortening slide preparation time. Through a deep learning model trained on a large dataset of pixel-level labeled terahertz images, it achieves automatic and accurate segmentation and identification of cancer cell regions. This model employs a convolutional neural network with an encoder-decoder structure, optimizing boundary learning through a composite loss function to output high-precision cancer cell localization and contour information. Furthermore, this invention integrates terahertz near-field imaging hardware with artificial intelligence analysis software, forming a complete intraoperative intelligent assisted diagnostic system. The system can automatically complete the entire process from slide scanning, image processing, AI analysis to report generation, and visually overlay the identified cancer region contours onto the original terahertz image, providing pathologists with objective and visualized decision support. This improves diagnostic accuracy and consistency while meeting the time requirements for rapid intraoperative diagnosis.

[0067] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for intelligent-assisted diagnosis of thyroid cancer using intraoperative frozen section, characterized in that, Includes the following steps: S1. Prepare intraoperative frozen sections of thyroid tissue to be diagnosed, wherein the frozen sections are not stained or mounted. S2. Scan the frozen section using a terahertz near-field imaging system to obtain terahertz near-field image data of the frozen section; S3. Input the terahertz near-field image data into a pre-trained thyroid cancer cell recognition model, and have the model output a segmentation result map corresponding to the cancer cell region in the frozen section; wherein, the thyroid cancer cell recognition model is obtained by training a deep learning network based on a training dataset containing a large number of terahertz near-field images of thyroid tissue and corresponding pixel-level cancer cell contour annotations. S4. Based on the segmentation result image, generate a visualization report to assist in intraoperative pathological diagnosis.

2. The method for intelligent assisted diagnosis of intraoperative frozen section in thyroid cancer according to claim 1, characterized in that, The terahertz near-field imaging system includes a terahertz source, a focusing lens group for focusing terahertz waves, a metal probe for near-field detection, a detector, and a signal processing and imaging module.

3. The method for intelligent assisted diagnosis of thyroid cancer via intraoperative frozen section as described in claim 1, characterized in that, The pixel-level cancer cell outlines in the dataset are generated based on the diagnostic results of the corresponding paraffin sections.

4. A method for intelligent assisted diagnosis of thyroid cancer via intraoperative frozen section as described in claim 1 or 3, characterized in that, The deep learning network is a convolutional neural network with an encoder-decoder structure, and the training dataset contains terahertz near-field images and their annotations of no less than 1,000 cases of thyroid cancer.

5. The method for intelligent assisted diagnosis of intraoperative frozen section in thyroid cancer according to claim 1, characterized in that, The convolutional neural network with an encoder-decoder structure is U-Net or a variant thereof.

6. A smart assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer for implementing the method of any one of claims 1-5, characterized in that, include: The terahertz near-field imaging module is used to scan frozen sections of thyroid tissue that have not been stained and mounted, and to acquire terahertz near-field image data at nanometer resolution. The data processing and analysis module is communicatively connected to the terahertz near-field imaging module and integrates a pre-trained thyroid cancer cell recognition model. The data processing and analysis module is used to receive the terahertz near-field image data, use the thyroid cancer cell recognition model to automatically identify and segment cancer cell regions in the image, and generate a visual diagnostic auxiliary report based on the segmentation results.

7. The intelligent assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer according to claim 6, characterized in that, The thyroid cancer cell recognition model was constructed in the following manner: A dataset consisting of multiple terahertz near-field images of thyroid tissue and their pixel-level cancer cell contour annotations was obtained; The dataset is used to train an initial deep learning network, which is used to perform image semantic segmentation tasks. When the training reaches the convergence condition, the thyroid cancer cell recognition model is obtained.

8. The intelligent assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer according to claim 7, characterized in that, The construction of the dataset includes: using annotation tools to outline the cancer cell region on the terahertz near-field image based on the gold standard diagnostic results of the corresponding paraffin sections, and generating an annotation mask that corresponds one-to-one with the pixels of the original image.

9. The intelligent assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer according to claim 1, characterized in that, The terahertz near-field imaging module is based on terahertz scattering scanning near-field optical microscopy technology using a metal probe tip.

10. The intelligent assisted diagnostic system for intraoperative frozen section analysis of thyroid cancer according to claim 1, characterized in that, The system also includes a display module for showing the pathologist the visual diagnostic aid report, which includes the result of overlaying the outline of the cancer region onto the original terahertz near-field image.

Images