Early tongue cancer latent lymph node metastasis prediction system, electronic equipment and storage medium

Abstract

The invention relates to an early tongue cancer latent lymph node metastasis prediction system, electronic equipment and a storage medium. The early tongue cancer latent lymph node metastasis prediction system comprises a data acquisition module used for acquiring head and neck MRI images, the head and neck MRI images being acquired MRI images of a target object; the automatic tumor detection model is used for performing target detection on the head and neck MRI image to obtain a tumor area image; the lymph node metastasis prediction model is used for performing lymph node metastasis prediction classification on the tumor area image to obtain a lymph node metastasis classification result; by means of the scheme, accurate tumor positioning of early oral tongue cancer can be achieved through the automatic tumor detection model, and accurate prediction of latent lymph node metastasis can be achieved through the lymph node metastasis prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic equipment, and in particular to a prediction system for occult lymph node metastasis of early tongue cancer, electronic equipment and storage medium. Background Art

[0002] Oral and tongue cancer is a common type of head and neck malignant tumor, 90% of which belong to squamous cell carcinoma. Regional lymph node metastasis (LNM), as a key factor affecting the prognosis of oral and tongue cancer, can significantly reduce the patient's expected survival rate by as much as 50%. It is particularly noteworthy that the proportion of occult LNM in early oral and tongue cancer is quite high, about 30%, which brings challenges to clinical diagnosis and treatment. Faced with the decision-making dilemma of cervical lymph node treatment in early oral cancer, the current guidelines have not yet formed a unified opinion, and the main strategies include selective neck dissection (END) and conservative observation. Although END has shown positive effects in reducing recurrence rate and improving survival rate in early oral and tongue cancer, its unnecessary surgery rate of up to 70%, additional medical costs and negative impact on patients' quality of life cannot be ignored. Therefore, accurate identification of LNM in early oral and tongue cancer is of great significance for guiding treatment decisions and optimizing patient prognosis.

[0003] In the field of preoperative evaluation, magnetic resonance imaging (MRI) is widely used as a non-invasive method, but its traditional methods have limitations in diagnosing LNM, such as false positives (especially in reactive and infected nodules) and inadequacies in detecting micrometastases. With the rise of deep learning technology, its application in the field of medical imaging has opened up new avenues for cancer diagnosis, prognosis assessment, and treatment planning. In particular, MRI-based deep learning technology has shown the potential to effectively identify LNM in a variety of cancers, such as breast cancer, cervical cancer, and rectal cancer. However, current MRI-deep learning research on oral LNM identification still faces challenges, including reliance on labeled images to build models, lack of cross-device verification, and insufficient generalization capabilities.

[0004] In view of this, it is particularly urgent to provide a technical solution for accurate tumor localization and accurate prediction of occult LNM in early oral and tongue cancer. Such technology can not only improve the accurate tumor localization and accurate prediction of occult LNM in early oral and tongue cancer, but also provide patients with more personalized and accurate treatment plans, thereby improving the overall treatment effect and the quality of life of patients. Summary of the invention

[0005] The purpose of the present application is to provide a prediction system for occult lymph node metastasis of early-stage tongue cancer, an electronic device and a storage medium, which can realize the precise tumor positioning and accurate prediction of occult LNM in early-stage oral tongue cancer.

[0006] According to a first aspect of an embodiment of the present application, a system for predicting occult lymph node metastasis of early tongue cancer is provided, the system comprising:

[0007] A data acquisition module, used for acquiring a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of a target object;

[0008] An automatic tumor detection model is used to perform target detection on the head and neck MRI image to obtain a tumor region image;

[0009] The lymph node metastasis prediction model is used to perform lymph node metastasis prediction classification on the tumor region image to obtain a lymph node metastasis classification result.

[0010] In one embodiment, the system further comprises: a nomogram construction module, configured to:

[0011] Stepwise multivariate logistic regression analysis was used to identify clinical predictors;

[0012] An automated deep learning nomogram was constructed based on the clinical predictors and the automated deep learning labels.

[0013] In one embodiment, the system further comprises: an image acquisition and annotation module and an automatic tumor detection model training module;

[0014] The image acquisition and annotation module is used to:

[0015] Collecting original head and neck MRI images, annotating the original head and neck MRI images, and obtaining head and neck MRI sample images;

[0016] The automatic tumor detection model training module is used to:

[0017] The initial deep learning model is trained according to the head and neck MRI sample images to obtain the automatic tumor detection model for performing target detection on the head and neck MRI images.

[0018] In one embodiment, the automatic tumor detection model training module, when training the initial deep learning model according to the head and neck MRI sample image to obtain the automatic tumor detection model for performing target detection on the head and neck MRI image, is used to:

[0019] The complete intersection-over-union ratio and the distributed focal loss are used as the loss function, and the head and neck MRI sample images are input into the initial deep learning model for training to obtain the automatic tumor detection model for target detection on the head and neck MRI images.

[0020] In one embodiment, the head and neck MRI sample images include an axial fat-suppressed T2-weighted imaging sequence and an axial fat-suppressed enhanced T1-weighted imaging sequence.

[0021] In one embodiment, the automatic tumor detection model includes: a feature extraction network and an output layer;

[0022] The feature extraction network is used to:

[0023] Extracting features from the head and neck MRI image to obtain high-dimensional features of the image;

[0024] The output layer is used to:

[0025] Target detection is performed based on the high-dimensional features of the image to obtain the tumor region image.

[0026] In one embodiment, the lymph node metastasis prediction model includes: a coding layer and a multi-layer perceptron layer;

[0027] The coding layer is used for:

[0028] Encoding the tumor region image to obtain deep learning features;

[0029] The multi-layer perceptron layer is used to:

[0030] Lymph node metastasis is predicted based on the deep learning features to obtain the lymph node metastasis classification result.

[0031] In one embodiment, the system further comprises: an image preprocessing module and a lymph node metastasis prediction model training module;

[0032] The image preprocessing module is used for:

[0033] Using a linear interpolation method, the resolution of the sample image of the tumor area is processed into set pixels; and

[0034] Cut out a central area sample image of a set size from the processed tumor area sample image;

[0035] The lymph node metastasis prediction model training module is used to:

[0036] The initial model is trained according to the central region sample image to obtain a lymph node metastasis prediction model for predicting and classifying lymph node metastasis of the tumor region image.

[0037] According to a second aspect of an embodiment of the present application, an electronic device is provided, comprising a memory and a processor, wherein the memory is used to store a computer program executable by the processor; and the processor is used to execute the computer program in the memory, so as to:

[0038] Acquire a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of the target object;

[0039] Performing target detection on the head and neck MRI image to obtain a tumor region image;

[0040] The tumor region image is subjected to lymph node metastasis prediction and classification to obtain a lymph node metastasis classification result.

[0041] According to a third aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the executable computer program in the storage medium is executed by a processor, it can be used to:

[0042] Acquire a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of the target object;

[0043] Performing target detection on the head and neck MRI image to obtain a tumor region image;

[0044] The tumor region image is subjected to lymph node metastasis prediction and classification to obtain a lymph node metastasis classification result.

[0045] Compared with the prior art, the beneficial effects of the present application are as follows: the early stage tongue cancer occult lymph node metastasis prediction system comprises: a data acquisition module, an automatic tumor detection model and a lymph node metastasis prediction model; the data acquisition module is used to acquire head and neck MRI images, wherein the head and neck MRI images are acquired MRI images of the target object; the automatic tumor detection model is used to perform target detection on the head and neck MRI images to obtain tumor area images; the lymph node metastasis prediction model is used to perform lymph node metastasis prediction and classification on the tumor area images to obtain lymph node metastasis classification results; through the above scheme, the automatic tumor detection model can realize the precise positioning of tumors in early stage oral tongue cancer, and the lymph node metastasis prediction model can realize the accurate prediction of occult LNM. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the structure of a system for predicting occult lymph node metastasis of early tongue cancer provided by an exemplary embodiment of the present application;

[0047] Figure 2 A workflow diagram for the development of a fully automated MRI deep learning framework provided for an exemplary embodiment of the present application;

[0048] Figure 3 A schematic diagram of an encoder ResNet50 provided for an exemplary embodiment of the present application;

[0049] Figure 4A schematic diagram of an encoder DINOv2 provided for an exemplary embodiment of the present application;

[0050] Figure 5 Schematic diagram of a representative case of early-stage tongue cancer in which lymph node information is correctly diagnosed by a fully automated deep learning framework provided by an exemplary embodiment of the present disclosure;

[0051] Figure 6 A schematic diagram of an automated deep learning nomogram constructed for an exemplary embodiment of the present application;

[0052] Figure 7 A schematic diagram comparing the areas under the curves of deep learning models of different cohorts provided for exemplary embodiments of the present disclosure;

[0053] Figure 8 A schematic diagram showing the performance comparison of automated deep learning nomograms and clinical T staging for different cohorts provided for exemplary embodiments of the present disclosure;

[0054] Fig. 9 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0055] Unless otherwise defined, the technical terms or scientific terms used in this specification and claims shall have the usual meanings understood by persons with ordinary skills in the technical field to which the invention belongs. The specific embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be noted that in the specific description of these embodiments, in order to provide a concise description, it is impossible for this specification to provide a detailed description of all the features of the actual embodiments. Without departing from the spirit and scope of the present invention, those skilled in the art may modify and replace the embodiments of the present invention, and the resulting embodiments are also within the scope of protection of the present invention.

[0056] Magnetic Resonance Imaging (MRI): is an advanced medical imaging technology that uses magnetic fields and radio waves to generate detailed images of the internal structure of the human body. Magnetic resonance imaging is a non-invasive medical imaging examination technology. Its basic principle is to use the magnetic resonance phenomenon that occurs when hydrogen nuclei in human tissues in a static magnetic field are excited by radio frequency pulses of a specific frequency. When the radio frequency pulse ends, the hydrogen nuclei will induce MR signals during the relaxation process. After the MR signal is received, spatially encoded, and image reconstructed, an MR image can be generated.

[0057] In order to solve the above-mentioned technical problems, in some embodiments of the present application, an early tongue cancer occult lymph node metastasis prediction system includes: a data acquisition module, an automatic tumor detection model and a lymph node metastasis prediction model; the data acquisition module is used to acquire head and neck MRI images, wherein the head and neck MRI images are acquired MRI images of the target object; the tumor automatic detection model is used to perform target detection on the head and neck MRI images to obtain tumor area images; the lymph node metastasis prediction model is used to perform lymph node metastasis prediction and classification on the tumor area images to obtain lymph node metastasis classification results; through the above scheme, the automatic tumor detection model can achieve accurate tumor positioning of early oral tongue cancer, and the lymph node metastasis prediction model can achieve accurate prediction of occult LNM.

[0058] Figure 1 The schematic diagram of the structure of a prediction system for occult lymph node metastasis of early tongue cancer provided by the exemplary embodiment of the present application. Figure 1 As shown, the system includes:

[0059] A data acquisition module, used for acquiring a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of a target object;

[0060] Automatic tumor detection model, used to detect targets in head and neck MRI images and obtain tumor area images;

[0061] The lymph node metastasis prediction model is used to predict and classify lymph node metastasis in tumor area images to obtain lymph node metastasis classification results.

[0062] The early stage tongue cancer occult lymph node metastasis prediction system comprises: a data acquisition module, an automatic tumor detection model and a lymph node metastasis prediction model; the data acquisition module is used to acquire head and neck MRI images, wherein the head and neck MRI images are acquired MRI images of target objects; the automatic tumor detection model is used to perform target detection on the head and neck MRI images to obtain tumor area images; the lymph node metastasis prediction model is used to perform lymph node metastasis prediction and classification on the tumor area images to obtain lymph node metastasis classification results; through the above scheme, the automatic tumor detection model can realize the precise positioning of tumors in early stage oral tongue cancer, and the lymph node metastasis prediction model can realize the accurate prediction of occult LNM.

[0063] The execution subject of the above-mentioned early tongue cancer occult lymph node metastasis prediction system of the present application can be a server. Among them, the implementation form of the server is not limited. For example, the server can be a server device such as a conventional server, a cloud server, a cloud host, a virtual center, etc. Among them, the composition of the server mainly includes a processor, a hard disk, a memory, a system bus, etc., and a general computer architecture type.

[0064] It should be noted that the target object in this application refers to an ordinary user, or it can also be an animal, that is, this application can predict tongue cancer lymph node metastasis for ordinary users, and can also predict tongue cancer lymph node metastasis for animals.

[0065] Accordingly, an MRI image of the head and neck of the target object may be acquired by an MRI acquisition device.

[0066] In addition, before using the automatic tumor detection model, the initial deep learning model needs to be trained to obtain the automatic tumor detection model. The automatic tumor detection model can automatically detect targets on head and neck MRI images, obtain tumor area images, and achieve accurate tumor localization of cavity tongue cancer.

[0067] In some embodiments of the present application, the early tongue cancer occult lymph node metastasis prediction system also includes: an image acquisition and annotation module and an automatic tumor detection model training module. Among them, the image acquisition and annotation module is used to acquire original head and neck MRI images, annotate the original head and neck MRI images, and obtain head and neck MRI sample images. The automatic tumor detection model training module trains the initial deep learning model according to the head and neck MRI sample images to obtain an automatic tumor detection model for target detection of head and neck MRI images. The present application performs data acquisition and annotation through the image acquisition and annotation module, and performs model training through the automatic tumor detection model training module to obtain an automatic tumor detection model. The specific implementation methods of the image acquisition and annotation module and the automatic tumor detection model training module are described below.

[0068] It should be noted that the head and neck MRI sample images include axial fat-suppressed T2-weighted imaging sequences and axial fat-suppressed enhanced T1-weighted imaging sequences.

[0069] In an exemplary embodiment, the image acquisition and annotation module performs MRT image acquisition. For example, MRI examinations were performed in two participating centers, using a total of nine high-end scanners from four internationally renowned suppliers, including six 3.0-T high-field scanners and three 1.5-T scanners. These devices include the Ingenia, Ingenia CX and Achieva series of Philips Healthcare of the Netherlands, the Magnetom Vida and Magnetom Prisma series of Siemens Healthineers of Germany, the Signa Premier, Signa Excite and Signa HDxt series of GE Medical Systems of the United States, and the uMR560 of Shanghai UI Healthcare, a local Chinese brand. To ensure the accuracy of image analysis, the study selected axial fat-suppressed T2-weighted imaging sequences (T2WI) and axial fat-suppressed enhanced T1-weighted imaging sequences (ceT1WI).

[0070] In another exemplary embodiment, the image acquisition and annotation module performs image annotation. For example, in the image post-processing stage, the study used ITK-SNAP software to accurately delineate the tumor area. In the specific operation, the region of interest (ROI) is carefully depicted by stacking layer by layer to ensure comprehensive and accurate coverage. First, a radiologist with extensive experience in head and neck MRI diagnosis performed preliminary manual annotation work without referring to any pathological results or other imaging reports. Subsequently, these annotation results were reviewed by another senior radiologist with extensive experience in head and neck MRI diagnosis to ensure that the outline of each tumor was accurately outlined. This process strictly follows the principle of blind evaluation, aiming to minimize the impact of subjective bias on research results and to improve the accuracy of image annotation.

[0071] It should be noted that the automatic tumor detection model of this application adopts the YOLOv8 model. In the training process of the automatic tumor detection model, the complete intersection-over-union ratio and the distribution focal loss are used as the loss function.

[0072] In another exemplary embodiment, the automatic tumor detection model training module trains the initial deep learning model based on the head and neck MRI sample images to obtain an automatic tumor detection model for target detection in head and neck MRI images. For example, all image data of the two imaging sequences of ceT1WI and T2WI are fully utilized, regardless of whether they contain tumor areas, and are included in the training data set of the YOLOv8 model. This approach ensures that the model can learn rich image features and improve its detection capabilities in complex scenarios. At the same time, the data set also contains images of tumor areas carefully outlined by two senior radiologists, providing high-quality supervision information for the model.

[0073] In an exemplary embodiment, the automatic tumor detection model training module trains the initial deep learning model based on the head and neck MRI sample images to obtain the automatic tumor detection model for target detection in the head and neck MRI images. An achievable method is to use complete intersection-over-union and distributed focal loss as loss functions, input the head and neck MRI sample images into the initial deep learning model for training, and obtain the automatic tumor detection model for target detection in the head and neck MRI images.

[0074] In the above embodiment, in order to further improve the detection performance, two advanced bounding box loss functions are used: complete intersection over union (CIoU) and distribution focus loss (DFL). As an improved version of IoU, CIoU not only considers the overlapping area between bounding boxes, but also introduces geometric factors such as distance and aspect ratio, so that the match between the predicted bounding box and the true bounding box is more accurate. DFL optimizes the loss distribution, so that the model pays more attention to samples that are difficult to classify, thereby improving the robustness of the overall detection.

[0075] It should be noted that the automatic tumor detection model includes: a feature extraction network and an output layer. The specific implementation of the feature extraction network and the output layer is described below.

[0076] After the automatic tumor detection module is trained, the head and neck MRI images are input into the automatic tumor detection model, and the automatic tumor detection model performs target detection on the head and neck MRI images to obtain tumor area images. Among them, the head and neck MRI images are input into the automatic tumor detection model, and the feature extraction network is used to extract features from the head and neck MRI images to obtain high-dimensional features of the image; the output layer is used to perform target detection based on the high-dimensional features of the image to obtain the tumor area image.

[0077] Figure 2 A workflow diagram for developing a fully automatic MRI deep learning framework is provided for an exemplary embodiment of the present application. Figure 2As shown in the figure, the process mainly includes: MRI acquisition and image annotation, deep learning for automatic tumor detection, deep learning for lymph node metastasis prediction, and the development of automatic deep learning nomograms. Among them, the core of this process is to use an advanced deep learning model, You Only Look Once version8 (YOLOv8), which is specifically used for tumor detection tasks. The network architecture of YOLOv8 is exquisitely designed. Its core is the optimized CSPDarknet53 backbone network, which incorporates a cross-stage partial bottleneck structure and is equipped with two convolutional layers. It aims to significantly improve the accuracy of tumor detection by efficiently fusing high-level features and contextual information. In the output layer, YOLOv8 cleverly uses the Sigmoid function as an activation mechanism to calculate the probability of containing the target object in the bounding box, thereby achieving accurate identification of potential tumor areas. In addition, the model also integrates the anchor-free design concept and the spatial pyramid pooling fast (SPPF) layer. These innovations not only enhance the generalization ability of the model, but also enable it to be competent for multi-scale prediction tasks, further improving the comprehensiveness and accuracy of detection.

[0078] In addition, this application uses mean average precision (mAP) as the evaluation metric, which provides a comprehensive and objective evaluation of the overall detection performance of the model by comprehensively considering the precision and recall rate under different object categories and IoU thresholds.

[0079] In some embodiments of the present application, before using the lymph node metastasis prediction model, it is necessary to first train the lymph node metastasis prediction model.

[0080] The early tongue cancer occult lymph node metastasis prediction system also includes: an image preprocessing module and a lymph node metastasis prediction model training module. The image preprocessing module is used to process the resolution of the tumor area sample image to a set pixel using a linear interpolation method; and to crop a central area sample image of a set size from the processed tumor area sample image. The lymph node metastasis prediction model training module is used to train the initial model based on the central area sample image to obtain a lymph node metastasis prediction model for predicting and classifying lymph node metastasis in the tumor area image.

[0081] It should be noted that the present application does not limit the type of linear interpolation method, and the linear interpolation method includes but is not limited to: bilinear interpolation or cubic interpolation and other advanced interpolation methods to reduce pixel distortion and ensure image quality.

[0082] The set pixel may be a pixel value set according to an actual application scenario. The set pixel may be, for example, 448×448 pixels. The present application uses a linear interpolation method to standardize the resolution of the sample image of the tumor region to 448×448 pixels.

[0083] The size can be set according to the actual application scenario. For example, the size is set to a region of 55 mm in length and 35 mm in width.

[0084] For example, to ensure that the classification model can process input images of uniform specifications, a linear interpolation method was used to standardize the resolution of all images to 448×448 pixels. Label generation combines manual annotation with the automatic detection results of the YOLOv8 model, with special attention paid to and identification of image slices containing the largest tumor areas. Subsequently, areas with specific dimensions of 55 mm in length and 35 mm in width were cropped centered on these core areas. This precise cropping ensures that the prediction model can focus on the image features that are most closely related to tumor characteristics. In view of the possible limitations of the training data, a data augmentation strategy was implemented, which effectively quadrupled the size of the original training data set through a variety of technical means such as cropping, rotation, scaling, and flipping, thereby enhancing the generalization ability of the model.

[0085] After the lymph node metastasis prediction model is trained, the tumor region image is input into the lymph node metastasis prediction model, and the lymph node metastasis prediction model performs lymph node metastasis prediction classification on the tumor region image to obtain lymph node metastasis classification results. The lymph node metastasis classification results include two classification results: lymph node metastasis and lymph node metastasis.

[0086] In some embodiments of the present application, the lymph node metastasis prediction model includes: a coding layer and a multi-layer perceptron layer; the coding layer is used to encode the tumor area image to obtain deep learning features; the multi-layer perceptron layer is used to predict lymph node metastasis based on the deep learning features to obtain lymph node metastasis classification results. Among them, the coding layer uses two cutting-edge image encoders: ResNet50 based on convolutional neural network (CNN) and DINOv2 based on vision transformer (ViT). Among them, Figure 3 A schematic diagram of an encoder ResNet50 provided for an exemplary embodiment of the present application, Figure 4 A schematic diagram of an encoder DINOv2 provided in an exemplary embodiment of the present application.

[0087] Combination Figure 2-Figure 4In the feature extraction stage, this application uses two cutting-edge image encoders: ResNet50 based on convolutional neural network (CNN) and DINOv2 based on vision transformer (ViT). These two encoders have their own characteristics. ResNet50 is good at capturing local details of images by virtue of its supervised learning pre-training on large datasets such as ImageNet; while DINOv2 is pre-trained on a wider and more diverse unlabeled dataset using self-supervised learning, and can capture higher-level image representations. They work together on the tumor area image processed by YOLOv8 to extract deep learning features. Next, these features are sent to the multi-layer perceptron (MLP) layer for further processing. When the input image is a ceT1WI or T2WI sequence, the MLP layer will calculate and output the prediction scores based on each sequence respectively. If two sequences are input at the same time, their features are first extracted and connected respectively, and then comprehensively evaluated by the MLP layer to obtain the final prediction score. During this process, based on the prediction results of the model with the best automatic lesion detection performance, the automated deep learning signature was selected as the benchmark for subsequent analysis and application. The entire deep learning model architecture was built using the Python programming language, ensuring flexibility and efficiency in model development.

[0088] Figure 5 Schematic diagram of a representative case of early tongue cancer in which the fully automated deep learning framework provided by the exemplary embodiment of the present disclosure correctly diagnoses lymph node information. Figure 5 As shown, enhanced T1WI (rows a and c) and T2WI (rows b and d) images of patients without lymph node metastasis (LNM) (a and b) and patients with LNM (c and d) are shown. From left to right, each image sequence shows the original image, the bounding box based on manual annotation, the bounding box detected by YOLO, the cropped region of interest, and the heat map. The heat map shows that the key feature components obtained by reducing the features extracted by the ViT backbone network to five dimensions through principal component analysis are consistent with the tumor region, thus emphasizing the powerful feature extraction ability of the backbone network.

[0089] In some embodiments of the present application, the prediction system for occult lymph node metastasis of early tongue cancer further includes: a nomogram construction module. The nomogram construction module uses stepwise multivariate logistic regression analysis to determine clinical predictors; and constructs an automated deep learning nomogram based on clinical predictors and automatic deep learning labels.

[0090] For example, Figure 6 Schematic diagram of an automated deep learning nomogram constructed for an exemplary embodiment of the present application. Figure 6As shown, this application uses stepwise multivariate logistic regression analysis to identify independent predictors among clinical characteristics. Independent clinical predictors and automated deep learning labels were included in multivariate analysis to construct an automated deep learning nomogram. The Hosmer–Lemeshow test was used to evaluate the calibration of the automated deep learning nomogram. Decision curve analysis was used to quantify the net benefit of using clinical factors and automated deep learning nomograms at different threshold probabilities.

[0091] This application is a retrospective study, which has been formally approved by the Ethics Review Committee of Shanghai Ninth People's Hospital (SH9H, as center 1) and the First Affiliated Hospital of Guangxi Medical University (GMU1H, as center 2), and the need for written informed consent from patients was waived in accordance with regulations. The study focuses on patients with oral squamous cell carcinoma (specifically oral tongue cancer) who were treated at SH9H between January 2013 and June 2024 and at GMU1H ​​between January 2019 and January 2024.

[0092] Patients included in this study had to meet the following conditions: (1) completed enhanced head and neck magnetic resonance imaging (MRI) examination within 14 days before surgery; (2) preoperative clinical stage was cT1-2N0M0, that is, the primary tumor was limited to the early stage, without regional lymph node metastasis and distant metastasis; (3) had received primary tumor resection and selective neck lymph node dissection as a treatment plan; (4) had a complete lymph node pathology report. At the same time, patients with the following conditions were excluded: (1) the maximum axial diameter of the mass was less than 10 mm; (2) MRI image quality was poor or contained significant artifacts, affecting accurate diagnosis. To ensure the objectivity and professionalism of the evaluation, the study invited two experienced radiologists to participate. They had 9 and 14 years of profound attainments in the field of head and neck imaging, respectively. Without knowing the patient's histological results in advance, the two physicians independently evaluated the clinical N stage of all lymph nodes visible in the MRI images according to the established head and neck cancer LNM imaging evaluation criteria. Any disagreement in the evaluation was reached through consultation between the two parties.

[0093] Finally, 312 patients were screened and included in this study from the SH9H center, and 36 patients were included from the GMU1H ​​center. To further verify the performance of the model, patients in SH9H who were scanned with Philips or Siemens devices were randomly divided into a training group (n=163) and a validation group (n=65), while the remaining 84 patients who were scanned with GE or UI devices constituted an internal test group. In addition, all patients in GMU1H ​​who were scanned with GE, Philips, or Siemens devices (a total of 36 patients) served as an external test group to evaluate the cross-institutional generalization ability of the model.

[0094] In some embodiments of the present application, statistical analysis was performed using R software to evaluate the clinical feature differences between patients with and without LNM. According to the nature and distribution of the variables, all variables were compared using chi-square test or Mann–Whitney ∪ test. The diagnostic performance of clinical factors, deep learning models, and deep learning nomograms was evaluated using receiver operating characteristic (ROC) curve analysis. The area under the curve (AUC), accuracy, sensitivity, and specificity of all cohorts were calculated. AUC values ​​were compared using the DeLong method. Statistical significance was set at P<0.05.

[0095] Figure 7 Schematic diagram of the area under the curve of the deep learning model of different queues provided by the exemplary embodiment of the present disclosure. Figure 5 As shown in the figure, the bar graph shows the area under the receiver operating characteristic curve (AUC) of the deep learning model for predicting lymph node metastasis using manual annotation and automatic detection methods in the validation cohort (a), internal test cohort (b), and external test cohort (c).

[0096] Figure 8 Schematic diagram of the performance of automated deep learning nomograms and clinical T staging for different cohorts provided by the exemplary embodiment of the present disclosure. Figure 8 As shown, a, receiver operating characteristic (ROC) curve. b, confusion matrix of the automated deep learning nomogram. c, calibration curve of the automated deep learning nomogram. The calibration curve shows that there is a high degree of consistency between the probability of lymph node metastasis (LNM) predicted by the nomogram and the observed probability of LNM. d, decision curve analysis (DCA) curve shows that in most reasonable threshold probability ranges, the net benefit of the nomogram is higher than clinical T staging and simple diagnosis (such as all patients with or without LNM).

[0097] The performance of this application is described below in conjunction with experimental data.

[0098] Patient characteristics analysis. In the study involving two medical centers, 128 of 312 patients in center 1 were diagnosed with occult LNM, while 9 of 36 patients in center 2 were detected with LNM. Despite the difference in sample size, the statistical results showed that the incidence of LNM between the two centers was not significantly different (P = 0.092). Further analysis found that the LNM-positive and -negative groups showed significant differences in MRI deep invasion (DOI) and clinical T stage in the training, validation, internal, and external testing cohorts. Notably, the maximum tumor diameter showed significant between-group differences only in the training cohort. In particular, for the training cohort, we confirmed clinical T stage as an independent and powerful factor for predicting LNM by stepwise logistic regression analysis (odds ratio: 6.09; 95% confidence interval: 1.99–18.6, P = 0.002).

[0099]

[0100]

[0101] Table 1. Performance of YOLO v8 network in automatic tumor detection

[0102] Referring to Table 1 above, the performance of the YOLOv8 network on both ceT1WI and T2WI sequences is extremely outstanding. In the validation cohort, the mean average precision (mAP) of T1 and T2 tumors on ceT1WI reached 0.914 and 0.924, respectively, and on T2WI, it was 0.892 and 0.946, respectively. Similar high performance was also verified in the internal test cohort, with mAP of 0.877 and 0.899 for T1 and T2 tumors on ceT1WI, and 0.973 and 0.961 on T2WI. Even in the external test cohort, YOLOv8 maintained a stable performance, with mAP of 0.954 and 0.689 for T1 and T2 tumors on ceT1WI, and 0.897 and 0.895 on T2WI.

[0103]

[0104]

[0105] Table 2. Prediction performance of deep learning models based on manual annotations

[0106]

[0107] Table 3. Prediction performance of deep learning models based on automated segmentation Referring to Tables 2 and 3 above, overall, the ViT model outperformed the CNN model, and the ViT model that integrated the two sequences showed the best performance regardless of manual annotation or automatic detection (Table 2). Specifically, the highest AUC values ​​of the CNN model based on manual annotation in the validation cohort, internal test cohort, and external test cohort were 0.787, 0.606, and 0.691, respectively, while the highest AUC values ​​of the CNN model based on automatic detection in the same cohorts were 0.723, 0.573, and 0.691, respectively. In contrast, the AUC values ​​of the optimal ViT model based on manual annotation in the validation cohort, internal test cohort, and external test cohort were 0.835, 0.782, and 0.893, respectively, and the corresponding accuracies were 0.785, 0.750, and 0.833, respectively. Similarly, the AUC values ​​of the optimal model based on automatic detection in the same cohort were 0.856, 0.759, and 0.905, respectively, and the corresponding accuracies were 0.785, 0.679, and 0.806, respectively. There was no significant difference in the AUC values ​​between the optimal models based on manual annotation and automatic detection in the three cohorts (P = 0.558, 0.488, and 0.846). Finally, the ViT model based on automatic detection and integrating two sequences was selected as the automated deep learning signature.

[0108]

[0109] Table 4. Prediction performance of clinical T stage and automated DL nomogram

[0110] Referring to Table 4 above, the AUC values ​​of clinical T stage in the validation cohort, internal test cohort, and external test cohort were 0.653, 0.611, and 0.685, respectively, and the corresponding accuracies were 0.523, 0.548, and 0.528, respectively. In the same cohort, the AUC values ​​of the automated deep learning nomogram were 0.871, 0.776, and 0.909, respectively, and the corresponding accuracies were 0.785, 0.679, and 0.778, respectively. In the three cohorts, there was a significant difference between the AUC values ​​of the automated deep learning nomogram and the clinical T stage (P<0.001, P=0.002, and P<0.001). In the validation cohort, internal test cohort, and external test cohort, the automated deep learning nomogram showed good calibration ability, and the results of the Hosmer-Lemeshow test were not significant (P=0.980, 0.643, and 0.981). Finally, decision curve analysis showed that in the three cohorts, the automated deep learning nomogram achieved higher overall net benefits than clinical T stage within most reasonable threshold probability ranges.

[0111] The embodiment of the present application further provides an electronic device, comprising a processor and a memory; the memory is used to store a computer program executable by the processor; the processor is used to execute the computer program in the memory to implement:

[0112] A data acquisition module, used for acquiring a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of a target object;

[0113] Automatic tumor detection model, used to detect targets in head and neck MRI images and obtain tumor area images;

[0114] The lymph node metastasis prediction model is used to predict and classify lymph node metastasis in tumor area images to obtain lymph node metastasis classification results.

[0115] Optionally, the system further comprises: a nomogram building module, configured to:

[0116] Stepwise multivariate logistic regression analysis was used to identify clinical predictors;

[0117] An automated deep learning nomogram was constructed based on clinical predictors and automated deep learning labels.

[0118] Optionally, the system further comprises: an image acquisition and annotation module and an automatic tumor detection model training module;

[0119] Image acquisition and annotation module, used for:

[0120] Collecting original head and neck MRI images, annotating the original head and neck MRI images, and obtaining head and neck MRI sample images;

[0121] Automatic tumor detection model training module for:

[0122] The initial deep learning model is trained based on the head and neck MRI sample images to obtain an automatic tumor detection model for target detection in head and neck MRI images.

[0123] Optionally, the automatic tumor detection model training module, when training the initial deep learning model according to the head and neck MRI sample images to obtain the automatic tumor detection model for target detection on the head and neck MRI images, is used to:

[0124] Using complete intersection-over-union and distributed focal loss as loss functions, the head and neck MRI sample images were input into the initial deep learning model for training, and an automatic tumor detection model for target detection in head and neck MRI images was obtained.

[0125] Optionally, the head and neck MRI sample images include an axial fat-suppressed T2-weighted imaging sequence and an axial fat-suppressed enhanced T1-weighted imaging sequence.

[0126] Optionally, the automatic tumor detection model comprises: a feature extraction network and an output layer;

[0127] Feature extraction network for:

[0128] Perform feature extraction on head and neck MRI images to obtain high-dimensional features of the images;

[0129] Output layer, used to:

[0130] Target detection is performed based on the high-dimensional features of the image to obtain the tumor area image.

[0131] Optionally, the lymph node metastasis prediction model includes: an encoding layer and a multi-layer perceptron layer;

[0132] Encoding layer, used for:

[0133] Encode the tumor area image to obtain deep learning features;

[0134] Multilayer Perceptron layers, used to:

[0135] The deep learning features are used to predict lymph node metastasis and obtain the lymph node metastasis classification results.

[0136] Optionally, the system further comprises: an image preprocessing module and a lymph node metastasis prediction model training module;

[0137] Image preprocessing module for:

[0138] Using a linear interpolation method, the resolution of the sample image of the tumor area is processed into set pixels; and

[0139] Cut out a central area sample image of a set size from the processed tumor area sample image;

[0140] Lymph node metastasis prediction model training module, used for:

[0141] The initial model is trained based on the central area sample images to obtain a lymph node metastasis prediction model for predicting and classifying lymph node metastasis in tumor area images.

[0142] The embodiments of the present application further provide a computer-readable storage medium, which, when an executable computer program in the storage medium is executed by a processor, can be used to:

[0143] A data acquisition module, used for acquiring a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of a target object;

[0144] Automatic tumor detection model, used to detect targets in head and neck MRI images and obtain tumor area images;

[0145] The lymph node metastasis prediction model is used to predict and classify lymph node metastasis in tumor area images to obtain lymph node metastasis classification results.

[0146] Optionally, the system further comprises: a nomogram building module, configured to:

[0147] Stepwise multivariate logistic regression analysis was used to identify clinical predictors;

[0148] An automated deep learning nomogram was constructed based on clinical predictors and automated deep learning labels.

[0149] Optionally, the system further comprises: an image acquisition and annotation module and an automatic tumor detection model training module;

[0150] Image acquisition and annotation module, used for:

[0151] Collecting original head and neck MRI images, annotating the original head and neck MRI images, and obtaining head and neck MRI sample images;

[0152] Automatic tumor detection model training module for:

[0153] The initial deep learning model is trained based on the head and neck MRI sample images to obtain an automatic tumor detection model for target detection in head and neck MRI images.

[0154] Optionally, the automatic tumor detection model training module, when training the initial deep learning model according to the head and neck MRI sample images to obtain the automatic tumor detection model for target detection on the head and neck MRI images, is used to:

[0155] Using complete intersection-over-union and distributed focal loss as loss functions, the head and neck MRI sample images were input into the initial deep learning model for training, and an automatic tumor detection model for target detection in head and neck MRI images was obtained.

[0156] Optionally, the head and neck MRI sample images include an axial fat-suppressed T2-weighted imaging sequence and an axial fat-suppressed enhanced T1-weighted imaging sequence.

[0157] Optionally, the automatic tumor detection model comprises: a feature extraction network and an output layer;

[0158] Feature extraction network for:

[0159] Perform feature extraction on head and neck MRI images to obtain high-dimensional features of the images;

[0160] Output layer, used to:

[0161] Target detection is performed based on the high-dimensional features of the image to obtain the tumor area image.

[0162] Optionally, the lymph node metastasis prediction model includes: an encoding layer and a multi-layer perceptron layer;

[0163] Encoding layer, used for:

[0164] Encode the tumor area image to obtain deep learning features;

[0165] Multilayer Perceptron layers, used to:

[0166] The deep learning features are used to predict lymph node metastasis and obtain the lymph node metastasis classification results.

[0167] Optionally, the system further comprises: an image preprocessing module and a lymph node metastasis prediction model training module;

[0168] Image preprocessing module for:

[0169] Using a linear interpolation method, the resolution of the sample image of the tumor area is processed into set pixels; and

[0170] Cut out a central area sample image of a set size from the processed tumor area sample image;

[0171] Lymph node metastasis prediction model training module, used for:

[0172] The initial model is trained based on the central area sample images to obtain a lymph node metastasis prediction model for predicting and classifying lymph node metastasis in tumor area images.

[0173] Regarding the device in the above embodiment, the specific manner in which the processor performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0174] Fig. 9 900 is a block diagram of an electronic device according to an exemplary embodiment. For example, the electronic device 900 may be provided as a server. Fig. 9 The device 900 includes a processing component 922, which further includes one or more processors, and a memory resource represented by a memory 932 for storing instructions, such as applications, that can be executed by the processing component 922. The application stored in the memory 932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 922 is configured to execute instructions to perform the above-mentioned method for image processing.

[0175] The device 900 may also include a power component 926 configured to perform power management of the device 900, a wired or wireless network interface 950 configured to connect the device 900 to a network, and an input / output (I / O) interface 958. The device 900 may operate based on an operating system stored in the memory 932, such as Windows Server™, MacOS X™, Unix™, Linux™, FreeBSD™, or the like.

[0176] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 932 including instructions, which can be executed by the processing component 922 of the device 900 to perform the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0177] In the present invention, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless otherwise clearly defined.

[0178] The above description of the embodiments is to facilitate those of ordinary skill in the art to understand and apply the present application. It is obvious that those familiar with the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without having to pay creative labor. Therefore, the present application is not limited to the embodiments herein, and improvements and modifications made by those skilled in the art based on the contents disclosed in the present application without departing from the scope and spirit of the present application are within the scope of the present application.

Claims

1. A prediction system for occult lymph node metastasis of early tongue cancer, characterized in that: The system comprises: A data acquisition module, used for acquiring a head and neck MRI image, wherein the head and neck MRI image is an acquired MRI image of a target object; An automatic tumor detection model is used to perform target detection on the head and neck MRI image to obtain a tumor region image; The lymph node metastasis prediction model is used to perform lymph node metastasis prediction classification on the tumor region image to obtain a lymph node metastasis classification result.

2. The system according to claim 1, characterized in that The system further comprises: a nomogram building module, configured to: Stepwise multivariate logistic regression analysis was used to identify clinical predictors; An automated deep learning nomogram was constructed based on the clinical predictors and the automated deep learning labels.

3. The system according to claim 1, characterized in that The system also includes: an image acquisition and annotation module and an automatic tumor detection model training module; The image acquisition and annotation module is used to: Collecting original head and neck MRI images, annotating the original head and neck MRI images, and obtaining head and neck MRI sample images; The automatic tumor detection model training module is used to: The initial deep learning model is trained according to the head and neck MRI sample images to obtain the automatic tumor detection model for performing target detection on the head and neck MRI images.

4. The system according to claim 3, characterized in that The automatic tumor detection model training module, when training the initial deep learning model according to the head and neck MRI sample images to obtain the automatic tumor detection model for target detection on the head and neck MRI images, is used to: The complete intersection-over-union ratio and the distributed focal loss are used as the loss function, and the head and neck MRI sample images are input into the initial deep learning model for training to obtain the automatic tumor detection model for target detection on the head and neck MRI images.

5. The system according to claim 3, characterized in that The head and neck MRI sample images include an axial fat-suppressed T2-weighted imaging sequence and an axial fat-suppressed enhanced T1-weighted imaging sequence.

6. The system according to claim 1, characterized in that The automatic tumor detection model comprises: a feature extraction network and an output layer; The feature extraction network is used to: Extracting features from the head and neck MRI image to obtain high-dimensional features of the image; The output layer is used to: Target detection is performed based on the high-dimensional features of the image to obtain the tumor region image.

7. The system according to claim 1, characterized in that The lymph node metastasis prediction model comprises: a coding layer and a multi-layer perceptron layer; The coding layer is used for: Encoding the tumor region image to obtain deep learning features; The multi-layer perceptron layer is used to: Lymph node metastasis is predicted based on the deep learning features to obtain the lymph node metastasis classification result.

8. The system according to claim 1, characterized in that The system also includes: an image preprocessing module and a lymph node metastasis prediction model training module; The image preprocessing module is used to: Using a linear interpolation method, the resolution of the sample image of the tumor area is processed into set pixels; and Cut out a central area sample image of a set size from the processed tumor area sample image; The lymph node metastasis prediction model training module is used to: The initial model is trained according to the central region sample image to obtain a lymph node metastasis prediction model for predicting and classifying lymph node metastasis of the tumor region image.

9. An electronic device, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor for: Acquire MRI images of the head and neck, including: The head and neck MRI image is an acquired MRI image of the target object; Performing target detection on the head and neck MRI image to obtain a tumor region image; The tumor region image is subjected to lymph node metastasis prediction and classification to obtain a lymph node metastasis classification result.

10. A computer-readable medium having stored thereon at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by a processor for: Acquire MRI images of the head and neck, including: The head and neck MRI image is an acquired MRI image of the target object; Performing target detection on the head and neck MRI image to obtain a tumor region image; The tumor region image is subjected to lymph node metastasis prediction and classification to obtain a lymph node metastasis classification result.

Images