Solid tumor detection positioning system and method based on deep learning

Abstract

The invention relates to the technical field of medical image processing, and discloses a solid tumor detection positioning system and method based on deep learning, and the system comprises a data preprocessing module which is used for carrying out the registration, denoising and child age adaptive standardization processing of a child medical image; a deep learning segmentation model constructed based on a child solid tumor prior knowledge base is built in the tumor detection segmentation module, and the tumor detection segmentation module is used for outputting segmentation masks, categories and probabilities of tumors; the three-dimensional quantitative positioning module is used for performing three-dimensional reconstruction and spatial measurement on the segmentation result to generate a quantitative positioning report; and the collaborative iterative optimization module is used for man-machine interaction and model incremental learning. The method corresponds to the system. According to the method, the problems of high heterogeneity, data scarcity, insufficient quantitative analysis and model staticization of children tumors are solved, automatic and precise detection and positioning of the children solid tumors are realized, continuous optimization can be realized, and reliable assistance is provided for clinical diagnosis and treatment planning.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically a system and method for detecting and locating solid tumors based on deep learning. Background Technology

[0002] Early and accurate diagnosis of solid tumors in children is a significant clinical challenge. Compared to adults, pediatric tumors exhibit greater heterogeneity, and children are in a rapid growth and development phase, with organ size, proportions, and imaging characteristics changing significantly with age. This leads to more blurred lesion boundaries and a higher likelihood of missed diagnoses on CT and MRI images. Currently, clinical diagnosis relies heavily on the experience of radiologists, resulting in high subjectivity and low efficiency.

[0003] In recent years, although deep learning technology has been widely used in image analysis of adult tumors, its direct application to pediatric tumors has often yielded poor results. This is mainly because: First, existing general models do not consider the age-specific changes in pediatric images and lack effective standardized processing methods; second, high-quality pediatric tumor image data is scarce, making it difficult to directly train robust models; third, most studies only complete two-dimensional segmentation and fail to provide three-dimensional quantitative localization information that can be directly used for surgical or radiotherapy planning; and fourth, existing systems are usually static and lack the ability to self-optimize using clinical feedback.

[0004] Therefore, there is an urgent need for an intelligent assisted diagnostic solution that can be specifically designed for the characteristics of pediatric solid tumors, integrating automatic detection, precise segmentation, and three-dimensional quantitative analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based system and method for detecting and locating solid tumors, in order to solve the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention discloses the following technical solutions: In a first aspect, the present invention discloses a solid tumor detection and localization system based on deep learning, the system comprising: The data preprocessing module is configured to perform registration, denoising, and age-adaptive standardization processing on the input pediatric medical images to generate standardized 3D image data. The tumor detection and segmentation module is configured to process the standardized 3D image data using a built-in deep learning segmentation model, and output preliminary detection results with pixel-level tumor region segmentation masks, tumor categories, and tumor probabilities; wherein, the deep learning segmentation model is constructed and trained based on a fusion of a prior knowledge base of pediatric solid tumors; The three-dimensional quantitative localization module is configured to perform three-dimensional reconstruction of the tumor segmentation mask in the preliminary detection results, calculate the volume and centroid coordinates of the tumor, measure the minimum spatial distance between the tumor and preset surrounding key anatomical structures, and generate a three-dimensional quantitative localization report. The collaborative iterative optimization module is configured to provide an interactive interface to display the preliminary detection results and the three-dimensional quantitative localization report, receive correction feedback from physicians, and use the corrected data as new training samples to incrementally learn the deep learning segmentation model.

[0007] Optionally, the age-adaptive normalization process performed by the data preprocessing module includes: The input pediatric medical images are subjected to age-matched organ atlas registration, and the image space is normalized to a standard template; On the normalized images, the range of tumor-related regions of interest is adaptively determined based on the child's age and percentile of body surface area or height and weight, and the image intensity within the range is normalized by z-score.

[0008] Optionally, the construction and training process of the deep learning segmentation model includes: Model building steps: Construct a hybrid network based on encoder and decoder structure as the basic model, wherein the encoder uses a convolutional neural network to extract local features, and sets visual Transformer units in its bottleneck layer and deep layers of the decoder to capture global semantic context; Transfer learning steps: The base model is pre-trained using a large-scale adult abdominal solid tumor image dataset to obtain a pre-trained model; in the fine-tuning stage, structured sparse constraints related to pediatric tumor characteristics are applied to the pre-trained model based on the prior knowledge base of pediatric solid tumors to obtain a constrained model; the constrained model is fine-tuned using a pediatric solid tumor dataset to obtain the deep learning segmentation model.

[0009] Optionally, the tumor detection and segmentation module is configured to perform multi-task learning, the multi-task learning process including: Shared feature extraction step: Extract multi-level shared features of the standardized 3D image data through the encoder path shared by the encoder and decoder structures in the deep learning segmentation model; Segmentation task branch steps: Using the first task branch, perform pixel-level semantic segmentation based on the multi-level shared features to generate the tumor region segmentation mask; Classification task branch steps: Using the second task branch, based on the deep semantic features corresponding to the tumor region in the multi-level shared features, determine the tumor category and output the tumor category and the tumor probability.

[0010] Optionally, the loss function of the first task branch is a composite loss function for small lesions and blurred boundaries, which is expressed as follows: ,in For binary cross-entropy loss, For Dice similarity coefficient loss, To focus on the losses, , and To balance the weights.

[0011] Optionally, it also includes a model integration scheduling module, which includes: Multiple specific sub-models, each of which is a deep learning segmentation model optimized for a preset type of pediatric solid tumor; The model scheduler is used to select the corresponding specific sub-model to re-infer the standardized three-dimensional image data based on the tumor category with the highest probability in the preliminary detection results, and use the inference result as the final output.

[0012] Optionally, the three-dimensional quantitative localization module calculates the minimum spatial distance between the tumor and surrounding key anatomical structures by including: Based on the three-dimensional reconstructed surface of the tumor, calculate the shortest Euclidean distance from each vertex of its surface to a pre-defined triangular facet of the surrounding key anatomical structures. The minimum value among all the shortest Euclidean distances is taken as the minimum spatial distance between the tumor and the surrounding key anatomical structures.

[0013] Optionally, the incremental learning performed by the collaborative iterative optimization module includes: The tumor segmentation mask modified by the physician is used to form a new training sample pair with the corresponding pediatric medical images. The current deep learning segmentation model is further trained using the new training sample pairs and the historical training sample set at a smaller learning rate to iteratively update the model parameters.

[0014] Optionally, the pediatric medical images include computed tomography (CT) images and magnetic resonance imaging (MRI) images; the tumor categories defined in the pediatric solid tumor prior knowledge base include at least neuroblastoma, nephroblastoma, and hepatoblastoma.

[0015] Secondly, the present invention provides a method for detecting and locating solid tumors based on deep learning, the method comprising the following steps: The input pediatric medical images are registered, denoised, and standardized with age-adaptive processing to generate standardized 3D image data. The standardized 3D image data is processed by a deep learning segmentation model built and trained based on a prior knowledge base of pediatric solid tumors, and the preliminary detection results with pixel-level tumor region segmentation mask, tumor category and tumor probability are output. The tumor segmentation mask in the preliminary detection results is reconstructed in three dimensions, the volume and centroid coordinates of the tumor are calculated, and the minimum spatial distance between the tumor and the preset surrounding key anatomical structures is measured to generate a three-dimensional quantitative localization report. An interactive interface is provided to display the preliminary detection results and the three-dimensional quantitative localization report, receive correction feedback from physicians, and use the corrected data as new training samples to incrementally learn the deep learning segmentation model.

[0016] Beneficial Effects: The deep learning-based solid tumor detection and localization system and method of this invention, through age-adaptive standardization processing for children and a model construction training strategy that integrates a clinical prior knowledge base, enhances the adaptability of the deep learning segmentation model to the heterogeneity and imaging characteristics of pediatric tumors. This allows for high-sensitivity and specificity detection and segmentation even under data-scarce conditions. Furthermore, by generating a three-dimensional quantitative localization report containing information such as tumor volume and spatial distance, traditional image recognition is elevated to a quantitative analysis level that can directly assist clinical decision-making. In addition, incremental learning enables the model to continuously optimize in practical use, forming a virtuous cycle of increasing accuracy with use. Overall, this invention provides an automated, quantitative, and continuously evolving reliable tool for the early detection, accurate assessment, and treatment planning of pediatric solid tumors. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A structural block diagram of a deep learning-based solid tumor detection and localization system provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a deep learning-based method for detecting and locating solid tumors, as provided in an embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0020] Early and accurate diagnosis of pediatric solid tumors (such as neuroblastoma, nephroblastoma, and hepatoblastoma) is crucial for improving prognosis, but it faces more complex challenges than adult tumors. At the imaging level, four main interrelated technical difficulties limit the effectiveness of existing AI-assisted diagnostic systems.

[0021] First, there is the challenge of image standardization due to children's growth and development. Children are not simply miniature adults; their organ sizes, tissue proportions, and anatomical structures change significantly with age, height, and weight. This vast difference between and within individuals renders conventional medical image standardization methods based on fixed templates or simple linear scaling ineffective. If not handled properly, the same tumor will appear inconsistently in scale, location, and background contrast in images of different children, directly hindering subsequent deep learning models from learning stable and universal tumor features, severely impacting the model's generalization ability and detection robustness. How to achieve image preprocessing that adapts to children's physiological parameters (such as age and body surface area) is a primary prerequisite that must be addressed in constructing pediatric-specific models.

[0022] Second, there is the challenge of model training due to the scarcity of high-quality pediatric data and insufficient utilization of prior clinical knowledge. Solid tumors in children are relatively rare, resulting in an extreme shortage of high-quality, large-sample, and precisely labeled training datasets, making it difficult to train a high-performance deep learning model from scratch. While transfer learning is a common strategy to address data scarcity, directly using large-scale adult abdominal tumor data for pre-training pediatric models can introduce significant domain bias due to the inherent differences in the location, morphology, and biological behavior of tumors between adults and children. This can even lead to negative transfer, meaning the model learns interfering features that are not applicable to children. On the other hand, valuable structured prior knowledge accumulated by radiologists over long-term practice regarding the common sites of pediatric tumors and typical imaging features (such as "calcification" and "capsule") is difficult to effectively encode and utilize within existing end-to-end data-driven model building frameworks, resulting in knowledge waste.

[0023] Third, there is the challenge of clinical application due to the lack of information in 3D surgical planning, beyond the 2D pixel segmentation. Most existing research focuses on achieving pixel-level segmentation of tumor regions on a single or set of 2D slices. However, obtaining only a 2D outline is far from sufficient for surgical or radiotherapy planning. Clinically, there is an urgent need to know the precise 3D volume of the tumor to assess burden, its centroid spatial coordinates for localization, and the minimum invasive or compressive distance between its boundaries and nearby important blood vessels and organs to assess surgical risk and plan safety boundaries. A significant technological gap exists between 2D segmentation results and these directly guiding 3D quantitative parameters, which limits the clinical translational value of existing AI systems.

[0024] Fourth, there is the challenge of system iteration where static models are disconnected from dynamic clinical needs. Once deployed, traditional AI systems have fixed model parameters, becoming static systems. They cannot absorb feedback from experienced physicians addressing complex cases in real-world use, nor can they adapt to subtle differences in imaging equipment and diagnostic practices across different medical institutions. This "one-time training, permanent use" model prevents continuous system optimization, potentially leading to performance degradation when faced with out-of-distribution data or emerging atypical cases. It fails to grow in tandem with physicians' diagnostic skills, preventing the formation of a virtuous cycle of human-machine collaboration and increasing accuracy with use.

[0025] Finally, the assessment results are opaque and difficult to guide clinical action. Even when simple statistical models are introduced, their predictions often lack interpretability. Clinicians cannot understand why the model makes a certain risk judgment, nor are they clear about the key risk factors, leading to low trust in the assessment results and making it difficult to develop precise and personalized intervention strategies based on them.

[0026] In summary, existing technologies lack an integrated solution that can systematically address the four major challenges mentioned above. The embodiments of this application aim to overcome these deficiencies by providing a full-chain solution encompassing adaptive preprocessing, knowledge-guided model building, 3D quantitative localization, and human-machine collaborative iteration, with the goal of automating, refining, and continuously evolving the detection and localization of pediatric solid tumors.

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. Secondly, in this document, the term "comprising" is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0028] Furthermore, it should be clarified that the collection, processing, and use of all medical imaging data involved in the embodiments of this application are carried out in compliance with relevant laws, regulations, and ethical guidelines, and with the informed consent of the patient or their guardian and the authorization of the institution, to ensure that personal privacy and data security are fully protected.

[0029] Firstly, this embodiment provides a solid tumor detection and localization system based on deep learning, such as... Figure 1 As shown, the system includes: The data preprocessing module is configured to perform registration, denoising, and age-adaptive standardization processing on the input pediatric medical images to generate standardized 3D image data. The tumor detection and segmentation module is configured to process standardized 3D image data through a built-in deep learning segmentation model, and output preliminary detection results with pixel-level tumor region segmentation masks, tumor categories, and tumor probabilities; the deep learning segmentation model is built and trained based on a fusion of prior knowledge base of pediatric solid tumors; The three-dimensional quantitative localization module is configured to perform three-dimensional reconstruction of the tumor segmentation mask in the preliminary detection results, calculate the volume and centroid coordinates of the tumor, measure the minimum spatial distance between the tumor and the preset surrounding key anatomical structures, and generate a three-dimensional quantitative localization report. The collaborative iterative optimization module is configured to provide an interactive interface to display preliminary detection results and a 3D quantitative localization report, receive correction feedback from physicians, and use the corrected data as new training samples to incrementally learn the deep learning segmentation model.

[0030] In practice, the system is deployed on a server equipped with an NVIDIA A100 graphics card, connected to the hospital's Picture Archiving and Communication System (PACS) and physician workstations via a gigabit LAN. The data preprocessing module, tumor detection and segmentation module, 3D quantitative localization module, and collaborative iterative optimization module are four software functional units running on this server. They sequentially call and transmit data through predefined data interfaces (such as JSON files or memory objects in a specific format).

[0031] The data preprocessing module receives DICOM format computed tomography (CT) or magnetic resonance imaging (MRI) sequences of pediatric patients retrieved from the PACS system. The tumor detection and segmentation module encapsulates a pre-trained and deployed deep learning segmentation model; the model's construction and training process is detailed later. The 3D quantization and localization module utilizes algorithms from the open-source libraries VTK (Visualization Toolkit) and SimpleITK for 3D reconstruction and spatial computation. The collaborative iterative optimization module includes a visual interactive interface developed using web technologies (such as the Vue.js framework). This interface communicates with the backend server via the WebSocket protocol, receiving and displaying processing results in real time, while simultaneously encapsulating the doctor's corrections into standard labeled data formats (such as NIfTI format mask files) and sending them back to the server.

[0032] Based on the above, the data preprocessing module first normalizes the original images from diverse sources and with varying contrast to a standard space and intensity range using the age-adaptive method detailed in subsequent claims. This addresses the scale and contrast variations in pediatric images, providing stable and consistent input for subsequent models. The tumor detection and segmentation module utilizes a dedicated model incorporating prior clinical knowledge to reason about the standardized images, completing lesion localization (segmentation mask), identification (category), and confidence assessment (probability) in one go, solving the problem of poor model robustness under data scarcity. The 3D quantization and localization module transforms the 2D pixel-level segmentation results into volumetric data through a 3D reconstruction algorithm and accurately calculates key surgical parameters such as volume and spatial relationships, bridging the information gap between image recognition and clinical decision-making. The collaborative iterative optimization module incorporates the high judgment standard of doctors into the system closed loop. Through an incremental learning mechanism, the model can continuously optimize using clinical feedback, overcoming the performance stagnation of traditional static models. These four modules are interconnected, forming an automated, quantitative, and continuously evolving complete auxiliary diagnostic solution.

[0033] In this embodiment, pediatric medical images include computed tomography (CT) images and magnetic resonance imaging (MRI) images; the tumor categories defined in the pediatric solid tumor prior knowledge base include at least neuroblastoma, nephroblastoma, and hepatoblastoma.

[0034] Computed tomography (CT) images primarily refer to enhanced CT images, especially those in the venous phase. These images clearly show the tumor's blood supply, the presence or absence of calcification (important for neuroblastoma), and its relationship to major blood vessels. CT images are characterized by high resolution and fast scanning speed. Magnetic resonance imaging (MRI) images mainly refer to T1-weighted, T2-weighted, and diffusion-weighted imaging sequences. MRI provides excellent soft tissue contrast and is irreplaceable for assessing hepatoblastoma, determining internal tumor components (such as necrosis and hemorrhage), and observing spinal cord involvement (neuroblastoma). The system can receive DICOM format CT and MRI sequences. In practical applications, it can also be extended to support other modalities such as ultrasound imaging; this application does not limit this.

[0035] Furthermore, the prior knowledge base for pediatric solid tumors is a structured database or rule file, whose core encompasses at least the three most common pediatric malignant solid tumors: Neuroblastoma: It commonly occurs in the adrenal medulla or sympathetic nerve chain, and punctate calcifications are frequently seen on imaging. It is also prone to encircling blood vessels. Wilms' tumor: originates in the kidney, often presents as a large mass within the kidney, with claw sign (compression of remaining kidney tissue), and rarely invades renal blood vessels; Hepatoblastoma: Occurs in the liver, is most common in infants and young children, often shows significantly elevated AFP levels, and is usually a well-blooded mass on imaging.

[0036] The prior knowledge base for pediatric solid tumors defines fields for each type of tumor, such as its typical common sites (e.g., retroperitoneum, kidney, liver), common imaging features (e.g., calcification, enhancement pattern), and growth pattern (e.g., expansive, invasive).

[0037] Based on the above, by clearly defining the imaging modalities and tumor types, this system solution ensures that it addresses the specific clinical need for CT / MRI image analysis of pediatric abdominal solid tumors. The three listed tumor types are the main targets for system design, training, and optimization, guaranteeing the system's effectiveness and professionalism in this core task.

[0038] Simply scaling children's images to a fixed ratio or registering them using adult templates will result in severe distortion due to the dramatic non-linear changes in the size, shape, and relative position of children's organs with age, leading to the failure of subsequent model analysis. As an optional implementation in this embodiment, the data preprocessing module performs age-adaptive normalization processing for children, including: The input pediatric medical images are registered using age-matched organ atlases, and the image space is spatially normalized to a standard template. On the normalized images, the range of tumor-related regions of interest is adaptively determined based on the child's age and percentile of body surface area or height and weight, and the image intensity within the range is normalized by z-score.

[0039] In practical implementation, the system incorporates a probabilistic atlas of abdominal organs in children, categorized by age group. This atlas is built upon a large number of images of healthy children, including templates for five age ranges: 0-1 years, 1-3 years, 3-6 years, 6-12 years, and 12-18 years. The module first reads the patient's age information from the DICOM header file and selects the closest age group template. Then, an elastic registration algorithm (such as the SyN algorithm from the ANTs toolkit) is used to non-linearly register the input pediatric CT or MRI images with the selected template. This algorithm optimizes a deformation field to maximize the structural similarity between the input image and the template, thereby aligning the anatomical structures of all patients to a unified standard space. In practical applications, other registration toolkits such as Elastix can also be selected; this embodiment does not limit the choice.

[0040] In the spatially normalized images, the system no longer performs uniform processing on the entire image. Instead, it considers the patient's age. (age) and body surface area (m) 2Based on height and weight, a bounding box is determined as the region of interest (ROI) using a predefined rule. An example of the rule is as follows: For the retroperitoneal region (commonly found in neuroblastomas), the center point of its ROI... Relatively fixed in the template space, but its size ( For width, For height and (for depth) and They are positively correlated, and the calculation method is as follows: ,in These are the basic dimensions. This is a scaling factor, preset for different tumor-prone areas. This method ensures that the focus area of ​​the analysis covers the possible anatomical locations, regardless of the child's size.

[0041] Furthermore, image intensity normalization is performed only within the defined ROI. The mean intensity of all pixels within this ROI is calculated. and standard deviation Then, the intensity value of each pixel within the ROI is... Perform z-score normalization: Pixels outside the ROI are filled with a fixed value (such as 0). This method effectively eliminates intensity differences caused by different scanning devices and protocols, while avoiding the influence of background noise on the statistics.

[0042] Based on the above, the working principle of this processing method is to first spatially align, then locally focus, and finally correct the intensity. Age-specific template registration solves the core problem of the non-fixed spatial location of organs in children; adaptive ROI delineation based on physiological parameters focuses computational resources on tumor-prone areas, improving processing efficiency and reducing interference from irrelevant backgrounds; finally, z-score normalization within the ROI unifies the image contrast. These three steps work together to provide subsequent deep learning models with input data that is scale-uniform, roughly fixed in location, and has a standardized intensity distribution, reducing the difficulty of model learning and forming a fundamental step in improving the generalization performance of the entire system.

[0043] How to utilize massive amounts of adult data to compensate for insufficient pediatric data, while avoiding contamination or negative transfer of adult tumor features to pediatric models? As an optional implementation method in this embodiment, the construction and training process of the deep learning segmentation model includes: Model building steps: Construct a hybrid network based on encoder and decoder structure as the basic model. The encoder uses a convolutional neural network to extract local features, and sets visual Transformer units in its bottleneck layer and deep layers of the decoder to capture global semantic context. Transfer learning steps: The base model is pre-trained using a large-scale adult abdominal solid tumor image dataset to obtain a pre-trained model; in the fine-tuning stage, structured sparse constraints related to pediatric tumor characteristics are applied to the pre-trained model based on a prior knowledge base of pediatric solid tumors to obtain a constrained model; the constrained model is then fine-tuned using a pediatric solid tumor dataset to obtain a deep learning segmentation model.

[0044] In practice, the base model uses a 3D U-Net as its backbone. The encoder consists of five downsampling stages, each containing two 3×3×3 convolutional layers, a group normalization (GroupNorm) layer, and a LeakyReLU activation function, followed by a 2×2×2 max-pooling layer for downsampling. A visual Transformer unit (ViT) is connected to the last stage of the encoder (the bottleneck layer) and to the feature map of the deepest layer in the decoder. This ViT unit segments the input feature map into multiple 3D patches, flattens them, and obtains patch embeddings through linear projection, adding learnable positional encodings. These are then passed through multiple Transformer encoder layers (including multi-head self-attention mechanisms and feedforward networks) for global context modeling. The decoder performs upsampling through transposed convolutions and fuses the high-resolution features from the corresponding encoder layer with skip connections.

[0045] In pre-training, a publicly available dataset of CT images containing over 10,000 cases of adult kidney cancer, liver cancer, pancreatic cancer, etc., was used to pre-train the aforementioned base model. The learning task was semantic segmentation of tumor regions, and the loss function was a combination of Dice loss and cross-entropy loss. This step enabled the model to learn general low- and mid-level visual features for identifying and segmenting "tumor entities" from complex abdominal images.

[0046] Structured sparsity constraints are key to integrating prior knowledge. In this embodiment, the prior knowledge base for pediatric solid tumors is represented as a set of rules, such as: "Childhood neuroblastoma is often accompanied by punctate calcifications," "Hepatoblastoma is highly vascularized and shows significant enhancement," and "The tumor is closely related to surrounding large blood vessels (such as the inferior vena cava and abdominal aorta)." Before fine-tuning begins, the convolutional kernels of the pre-trained model are analyzed. Using a gradient saliency-based analysis method, the convolutional kernel channels that contribute the most to activating the aforementioned "adult-specific features" (such as typical liver cancer morphology) are identified. Then, based on the prior knowledge base, L1 regularization constraints are applied to these "adult-specific" channels, i.e., a regularized transport loss term is added to the loss function during the fine-tuning stage. .in, This represents the set of channel indices that need to be sparsified, selected based on prior knowledge. These are the convolution kernel weights for the corresponding channels. This is the constraint strength coefficient. During fine-tuning, this constraint causes the weights of these channels to tend towards zero, thereby suppressing the model's dependence on adult features to some extent.

[0047] During fine-tuning, a pre-trained model with sparse constraints was fine-tuned using a collected, desensitized, and labeled dataset of pediatric solid tumors (e.g., paired CT / MRI data containing 300 cases of neuroblastoma and 200 cases of nephroblastoma). At this point, while retaining general tumor segmentation capabilities, the model was forced to rely more heavily on the remaining channels to learn the unique features of pediatric tumors.

[0048] Based on the above, the core idea of ​​this construction and training process is to first perform general pre-training, followed by knowledge-guided specialization. The pre-training step fully leverages the scale advantage of adult data, enabling the model to master the basic capabilities of abdominal image analysis and tumor segmentation. The subsequent structured sparse constraints are not simply weight adjustments, but rather transform the prior clinical knowledge of differences between pediatric and adult tumors into direct guidance for the model's parameter space. This allows for the selective weakening of potentially interfering adult feature expression paths during fine-tuning, forcing the model to explore and strengthen feature channels more useful for pediatric tumor diagnosis. This approach solves the negative transfer problem when transferring learning from the adult domain to the pediatric domain, enabling the training of highly specialized and robust models even on limited pediatric data.

[0049] Using two separate models for processing would increase computational overhead and potentially lose inter-task correlation information. Therefore, as a further optional implementation of this embodiment, the tumor detection and segmentation module is configured to perform multi-task learning, the process of which includes: Shared feature extraction steps: Extract multi-level shared features from standardized 3D image data by using the encoder path shared by the encoder and decoder structures in the deep learning segmentation model; Segmentation task branch steps: Using the first task branch, perform pixel-level semantic segmentation based on multi-level shared features to generate a tumor region segmentation mask; Classification task branch steps: Using the second task branch, based on the deep semantic features corresponding to the tumor region in the multi-level shared features, the tumor category is determined, and the tumor category and tumor probability are output.

[0050] In practice, the encoder path (shared feature extraction step) of the model remains unchanged, responsible for extracting shared feature maps from multiple levels from low to high. .

[0051] The segmentation task branch serves as the main branch, inheriting the decoder structure from the original U-Net. The decoder processes the deepest features output by the encoder. Upsampling is performed and combined with the intermediate features of the encoder. Through skip connections and fusion, the final output, a tumor segmentation probability map of the same size as the input image, is obtained by passing the image through a 1×1×1 convolutional layer and a sigmoid activation function; this map serves as a tumor region segmentation mask. .

[0052] The classification task branch is an additional parallel branch. It first receives deep features from the encoder bottleneck layer, which contain rich global semantic information. Then, this branch uses a region-aware attention pooling module to process the initial segmentation mask generated in the previous step. (Or an intermediate segmentation result) is used as a soft attention weight for deep features. Perform weighted global average pooling (GAP). Specifically, the feature vector after weighted pooling... The calculation is as follows: ,in For spatial coordinates, To prevent division by zero for minimum values, this allows the classification branch to focus on features within the segmented region, rather than the entire image. Finally, The system uses a fully connected layer classifier (with the number of output units equal to the number of tumor categories, such as 3 categories) and a Softmax function to output the probability of each tumor category, i.e., the tumor category and the probability of the tumor.

[0053] Based on the above, the working mechanism of multi-task learning is feature sharing and task collaboration. The encoder, as a shared backbone network, extracts general features that simultaneously serve segmentation and classification, improving computational efficiency. The segmentation branch provides accurate spatial localization information (mask). The classification branch cleverly utilizes this localization information, automatically extracting the deep semantic information most relevant to the tumor region from shared features for classification through region attention convergence, avoiding interference from irrelevant background features. The two tasks are jointly optimized during training using a multi-task loss function, mutually reinforcing each other: better segmentation helps focus classification; more accurate classification semantic information can also indirectly guide the encoder to learn more discriminative features through gradient backpropagation, thereby improving segmentation accuracy. This design achieves integrated and efficient processing of detection (localization) and diagnosis (classification).

[0054] When segmenting branches, the small lesions and ambiguous boundaries of pediatric tumors can lead to severe class imbalance and training difficulties. Therefore, as a further optional implementation of this embodiment, the loss function of the first task branch is a composite loss function for small lesions and ambiguous boundaries, which is expressed as: ,in For binary cross-entropy loss, For Dice similarity coefficient loss, To focus on the losses, , and To balance the weights.

[0055] In practice, the aforementioned composite loss function is used when training the segmentation branch. Let the real label be... The model predicts the probability as , This represents the total number of pixels.

[0056] Binary cross-entropy loss is used to provide stable gradients, and its standard form is: .

[0057] Dice loss is used to directly optimize the overlap of segmented regions, and it is relatively sensitive to small targets. ,in This is a smoothing term.

[0058] Focal loss is used to address the imbalance between easy and difficult samples. It can be an adaptive modification of the standard Focal Loss, making it more focused on the tumor boundary region and the interior of small lesions that are difficult to segment, which are crucial to segmentation accuracy.

[0059] First, a basic Focal Loss term is defined to address the imbalance between positive and negative samples and to make the model focus more on pixels that are difficult to classify. Its basic form is: ,in, For focusing parameters ( ), used to reduce easily classifiable samples ( Loss weights that are close to 1 or 0. When At this point, the loss degenerates into the standard cross-entropy loss.

[0060] To explicitly guide the model to focus on the blurred region where tumors meet normal tissue, based on the true labels... Calculate a distance transformation graph . Each pixel value in the graph represents the Euclidean distance from that point to the nearest tumor boundary (points on the boundary have a distance of 0, which gradually increases towards the inside and outside of the tumor). A spatial weight map is then generated using a Gaussian decay function. .in, It is the boundary weight strength coefficient ( This is used to control the magnitude of loss enhancement in the boundary region. In practical applications, This can be determined by a grid search on the validation set, and the value typically ranges from [0.5, 2.0]. For example, in a specific configuration of this embodiment, the following settings are configured: . It is the boundary concern bandwidth coefficient ( This determines how far away from the boundary a pixel will receive a significant weight boost; its physical meaning is similar to the standard deviation of the Gaussian kernel. The settings are related to image resolution and can be initialized by analyzing the typical physical width of blurred tumor boundaries (e.g., a physical size corresponding to 1-2 pixels), and fine-tuned during training. In this embodiment, for Isotropic images, taking This weighted graph makes The pixels at the (boundary) level receive the highest weight. ,along with As the weighting factor increases, the weighting factor decreases to 1, which is equivalent to amplifying the loss contribution of the boundary pixels.

[0061] The final weighted focusing loss function incorporates boundary-aware weights. With basic Focal Loss items Multiplying these together yields the final improved focusing loss: During training, the and , According to the aforementioned weights Perform a weighted summation to construct the total loss function. .

[0062] Based on the above, this composite loss function works collaboratively through a dual modulation mechanism. First, the basic Focal Loss term... and Based on the model's predicted probability The confidence level is used to adaptively reduce the loss weights of easily classified samples, forcing the model to focus on optimizing those difficult pixels with uncertain predictions (usually complex textured areas inside lesions or parts of the background). Secondly, a boundary-aware weight map is used. It is a prior-driven static weight that does not rely on the model's instantaneous predictions. Instead, it directly assigns higher loss weight coefficients to boundary pixels based on anatomical priors (tumor boundaries are key and difficult-to-segment regions). This combination allows the loss function to adaptively focus on samples deemed difficult by the model while being explicitly guided to key regions (boundaries) determined by prior knowledge. This design is particularly suitable for scenarios with blurred and low-contrast tumor boundaries in children, effectively driving the model to invest more learning power in accurately delineating tumor edges and stabilizing the prediction of pixels within small lesions, thereby improving overall segmentation accuracy and boundary fit.

[0063] A single model may struggle to achieve optimal performance across all types of pediatric solid tumors. Neuroblastoma, nephroblastoma, and hepatoblastoma exhibit diverse imaging features and growth patterns, potentially leading to compromises in performance across certain types with a single, universal model. Therefore, as an optional implementation of this embodiment, a model integration scheduling module is also included, comprising: Multiple specific sub-models, each of which is a deep learning segmentation model optimized for a preset type of pediatric solid tumor; The model scheduler is used to select the corresponding specific sub-model to re-infer the standardized 3D image data based on the tumor category with the highest probability in the preliminary detection results, and use the inference result as the final output.

[0064] In practice, the model integration and scheduling module runs as an independent software service.

[0065] The specific sub-model library contains three deep learning segmentation models optimized for neuroblastoma, nephroblastoma, and hepatoblastoma, respectively. These sub-models all employ a similar structure to the aforementioned deep learning segmentation models, but are trained independently and thoroughly using their respective types of tumor data. For example, the neuroblastoma sub-model uses more training samples that include the retroperitoneal region and calcification features.

[0066] Secondly, the model scheduler is a lightweight decision logic unit. Its workflow is as follows: The system first processes the input image using the tumor detection and segmentation module to obtain preliminary detection results, including the most probable tumor category (e.g., nephroblastoma, probability 0.85) and its corresponding coarse segmentation mask. The scheduler reads the category label and then loads the corresponding specific sub-model (i.e., the nephroblastoma-specific model) from the sub-model library into GPU memory. Subsequently, this loaded specialized model is used to perform a second, more specialized inference calculation on the same standardized 3D image data. The typically more refined segmentation mask and category probability output by the specialized model will serve as the final output of this processing, used for subsequent 3D quantization and display. If the highest probability of the preliminary detection result is below a certain threshold (e.g., 0.5), the scheduler may determine it as uncertain and directly adopt the result of the general model or trigger manual review.

[0067] Based on the above, the model integration scheduling module adopts a two-stage strategy of coarse screening and fine-tuning. A general-purpose deep learning segmentation model plays the role of initial screening and triage; it is fast and has broad coverage, responsible for quickly locating suspected tumor regions and making preliminary classifications. Based on this preliminary classification result, the model scheduler intelligently calls specialized models that have been trained more thoroughly on specific types of tumors and have seen more variant samples for secondary analysis. The advantage of this design is that it maintains the system's compatibility with multiple tumor types (as guaranteed by the general-purpose model) while pursuing higher segmentation accuracy and classification confidence in specific cases by calling specialized models. This effectively balances the breadth and depth of the models, improving the stability and optimality of overall system performance.

[0068] For surgical planning, knowing only the shortest distance between the tumor and blood vessels is insufficient; it is also necessary to clarify which point on the tumor surface and which point on the blood vessel surface this distance represents. Therefore, as an optional implementation method in this embodiment, the calculation of the minimum spatial distance between the tumor and surrounding key anatomical structures by the three-dimensional quantitative localization module includes: Based on the three-dimensional reconstructed surface of the tumor, calculate the shortest Euclidean distance from each vertex of its surface to a pre-defined triangular patch of surrounding key anatomical structures. The minimum value among all the shortest Euclidean distances is taken as the minimum spatial distance between the tumor and the surrounding key anatomical structures.

[0069] In practice, the 3D quantization and localization module first performs surface meshing on the binary segmentation mask (3D volume data) of the tumor. Using the Marching Cubes algorithm, a 3D mesh model of the tumor surface is generated, consisting of numerous triangular facets (each facet containing 3 vertices). Similarly, for pre-defined key surrounding anatomical structures (such as the inferior vena cava, abdominal aorta, and kidneys), the system pre-loads or segments their surface mesh models from images of the same patient using similar methods. .

[0070] Calculate tumor surface each vertex to the anatomical surface Shortest spatial distance This is achieved through calculation. arrive All triangular facets Find the distance to the point and take the minimum value. To a triangular face (from vertex) The distance calculation (as defined) involves determining whether the projection of a point onto the face lies inside the triangle: if inside, the distance is the distance from the point to the plane; if outside, the distances to the three side segments are calculated. The specific calculation can be performed using an efficient geometry library (such as CGAL). This process can be formalized as: .

[0071] Finally, traverse all of the tumor surface Given vertices, obtain a set of distances. The minimum spatial distance between the tumor and this anatomical structure. That is: The system also records the acquisition. The coordinates of the tumor surface vertices and the corresponding anatomical structure surface points are used to highlight them in the visualization report.

[0072] Based on the above, this method works by precisely calculating surface geometry. Instead of approximating tumors and organs as spheres or ellipsoids to calculate center distances, it calculates the minimum distance from a set of surface points to a set of target facets, thus accurately reflecting the crucial geometric relationship of the shortest gap between two three-dimensional object surfaces. This method yields the minimum spatial distance... It is objective, accurate, and repeatable, providing doctors with a reliable and quantitative basis for determining whether a tumor has invaded blood vessels and for planning surgical margins, thus enhancing the clinical practical value of the system's output results.

[0073] Simply mixing new and old data for full retraining would consume enormous resources and could lead to data forgetting. Therefore, as a further optional implementation in this embodiment, the incremental learning performed by the collaborative iterative optimization module includes: The physician-corrected tumor segmentation mask is paired with the corresponding pediatric medical images to form a new training sample pair. With a small learning rate, the current deep learning segmentation model is continued to be trained using new training sample pairs and historical training sample sets to iteratively update the model parameters.

[0074] In practice, the system maintains a fixed-capacity historical training sample buffer (e.g., using a FIFO queue to store the most recent 1000 training samples). When a doctor corrects the segmentation result of a case through the interactive interface, the system uses the original image of that case and the corrected high-quality mask as a new training sample. Store in the buffer.

[0075] Every so often (e.g., accumulating 50 new samples) or periodically (e.g., weekly), the system automatically initiates an incremental training task. During training, instead of using the entire original dataset, a batch of samples (e.g., 128) is randomly sampled from the historical buffer and mixed with the latest batch of corrected samples (e.g., 32) to form a small training batch. A significantly reduced initial learning rate (e.g., ...) is used. The model in the current service is then trained using the aforementioned composite loss function (which is one-tenth of the original training data). The training cycle is relatively short (e.g., 5-10 rounds) to prevent overfitting to small batches of new data.

[0076] In addition, each incremental training iteration generates a new version of the model. Before switching to the new model, the system evaluates its performance on a separate validation set (composed of historical, uncorrected difficult cases). If the new model's key metrics (such as the average Dice coefficient) on the validation set drop by more than a preset threshold (e.g., 3%), it is determined that severe forgetting may have occurred, and the system will automatically roll back to the previous stable version and issue an alert to the administrator to check the quality of the new samples.

[0077] Based on the above, by using a very small learning rate and short-cycle training with batches of mixed new and old data, the updates to model parameters become very gentle and gradual. Gradients from new samples slightly adjust the model weights, adapting them to new annotation styles or allowing them to learn new case features; meanwhile, the gradients generated by the large number of historical samples act as anchors, continuously pulling the model back to its original knowledge space, effectively mitigating catastrophic forgetting. This history buffer mechanism ensures that the samples used for memory consolidation are dynamically updated and representative. This design allows the system to learn safely and continuously from doctors' daily feedback without costly retraining, achieving synchronous growth in model performance and diagnostic experience, forming an intelligent system that becomes more accurate with use.

[0078] Secondly, this embodiment provides a method for detecting and locating solid tumors based on deep learning, as described above (i.e., a method for detecting and locating solid tumors based on deep learning), such as... Figure 2 As shown, the method includes the following steps in sequence: S1 performs registration, denoising, and age-adaptive standardization processing on the input pediatric medical images to generate standardized 3D image data. S2 - A deep learning segmentation model, built and trained based on a prior knowledge base of pediatric solid tumors, processes standardized 3D image data and outputs preliminary detection results with pixel-level tumor region segmentation masks, tumor categories, and tumor probabilities. S3 - Performs three-dimensional reconstruction of the tumor segmentation mask in the preliminary detection results, calculates the tumor volume and centroid coordinates, measures the minimum spatial distance between the tumor and the preset surrounding key anatomical structures, and generates a three-dimensional quantitative localization report. S4 provides an interactive interface to display preliminary detection results and 3D quantitative localization reports, receives correction feedback from physicians, and uses the corrected data as new training samples to incrementally learn the deep learning segmentation model.

[0079] It should be noted that this method corresponds to the aforementioned deep learning-based solid tumor detection and localization system. Therefore, the parts of this method that are not specifically described (including but not limited to specific technical means and effects) can be referred to the relevant descriptions in the aforementioned deep learning-based solid tumor detection and localization system, and will not be repeated here.

[0080] In the embodiments provided by this invention, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor can be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments can be performed by a computer program instructing the associated hardware. During implementation, the program can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. Computer-readable storage media can include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0081] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based system for detecting and locating solid tumors, characterized in that, The system includes: The data preprocessing module is configured to perform registration, denoising, and age-adaptive standardization processing on the input pediatric medical images to generate standardized 3D image data. The tumor detection and segmentation module is configured to process the standardized 3D image data using a built-in deep learning segmentation model, and output preliminary detection results with pixel-level tumor region segmentation masks, tumor categories, and tumor probabilities; wherein, the deep learning segmentation model is constructed and trained based on a fusion of a prior knowledge base of pediatric solid tumors; The three-dimensional quantitative localization module is configured to perform three-dimensional reconstruction of the tumor segmentation mask in the preliminary detection results, calculate the volume and centroid coordinates of the tumor, measure the minimum spatial distance between the tumor and preset surrounding key anatomical structures, and generate a three-dimensional quantitative localization report. The collaborative iterative optimization module is configured to provide an interactive interface to display the preliminary detection results and the three-dimensional quantitative localization report, receive correction feedback from physicians, and use the corrected data as new training samples to incrementally learn the deep learning segmentation model.

2. The deep learning-based solid tumor detection and localization system according to claim 1, characterized in that, The data preprocessing module performs age-adaptive standardization processing for children, including: The input pediatric medical images are subjected to age-matched organ atlas registration, and the image space is normalized to a standard template; On the normalized images, the range of tumor-related regions of interest is adaptively determined based on the child's age and percentile of body surface area or height and weight, and the image intensity within the range is normalized by z-score.

3. The deep learning-based solid tumor detection and localization system according to claim 1, characterized in that, The construction and training process of the deep learning segmentation model includes: Model building steps: Construct a hybrid network based on encoder and decoder structure as the basic model, wherein the encoder uses a convolutional neural network to extract local features, and sets visual Transformer units in its bottleneck layer and deep layers of the decoder to capture global semantic context; Transfer learning steps: The base model is pre-trained using a large-scale adult abdominal solid tumor image dataset to obtain a pre-trained model; in the fine-tuning stage, structured sparse constraints related to pediatric tumor characteristics are applied to the pre-trained model based on the prior knowledge base of pediatric solid tumors to obtain a constrained model; the constrained model is fine-tuned using a pediatric solid tumor dataset to obtain the deep learning segmentation model.

4. The deep learning-based solid tumor detection and localization system according to claim 3, characterized in that, The tumor detection and segmentation module is configured to perform multi-task learning, the multi-task learning process including: Shared feature extraction step: Extract multi-level shared features of the standardized 3D image data through the encoder path shared by the encoder and decoder structures in the deep learning segmentation model; Segmentation task branch steps: Using the first task branch, perform pixel-level semantic segmentation based on the multi-level shared features to generate the tumor region segmentation mask; Classification task branch steps: Using the second task branch, based on the deep semantic features corresponding to the tumor region in the multi-level shared features, determine the tumor category and output the tumor category and the tumor probability.

5. The deep learning-based solid tumor detection and localization system according to claim 4, characterized in that, The loss function of the first task branch is a composite loss function for small lesions and blurred boundaries, which is expressed as follows: ,in For binary cross-entropy loss, For Dice similarity coefficient loss, To focus on the losses, , and To balance the weights.

6. The deep learning-based solid tumor detection and localization system according to claim 1, characterized in that, It also includes a model integration scheduling module, which comprises: Multiple specific sub-models, each of which is a deep learning segmentation model optimized for a preset type of pediatric solid tumor; The model scheduler is used to select the corresponding specific sub-model to re-infer the standardized three-dimensional image data based on the tumor category with the highest probability in the preliminary detection results, and use the inference result as the final output.

7. The deep learning-based solid tumor detection and localization system according to claim 1, characterized in that, The three-dimensional quantitative localization module calculates the minimum spatial distance between the tumor and surrounding key anatomical structures, including: Based on the three-dimensional reconstructed surface of the tumor, calculate the shortest Euclidean distance from each vertex of its surface to a pre-defined triangular facet of the surrounding key anatomical structures. The minimum value among all the shortest Euclidean distances is taken as the minimum spatial distance between the tumor and the surrounding key anatomical structures.

8. The deep learning-based solid tumor detection and localization system according to claim 7, characterized in that, The incremental learning performed by the collaborative iterative optimization module includes: The tumor segmentation mask modified by the physician is used to form a new training sample pair with the corresponding pediatric medical images. The current deep learning segmentation model is further trained using the new training sample pairs and the historical training sample set at a smaller learning rate to iteratively update the model parameters.

9. The deep learning-based solid tumor detection and localization system according to claim 1, characterized in that, The pediatric medical images include computed tomography (CT) images and magnetic resonance imaging (MRI) images; the tumor categories defined in the pediatric solid tumor prior knowledge base include at least neuroblastoma, nephroblastoma, and hepatoblastoma.

10. A method for detecting and locating solid tumors based on deep learning, characterized in that, The method includes the following steps: The input pediatric medical images are registered, denoised, and standardized with age-adaptive processing to generate standardized 3D image data. The standardized 3D image data is processed by a deep learning segmentation model built and trained based on a prior knowledge base of pediatric solid tumors, and the preliminary detection results with pixel-level tumor region segmentation mask, tumor category and tumor probability are output. The tumor segmentation mask in the preliminary detection results is reconstructed in three dimensions, the volume and centroid coordinates of the tumor are calculated, and the minimum spatial distance between the tumor and the preset surrounding key anatomical structures is measured to generate a three-dimensional quantitative localization report. An interactive interface is provided to display the preliminary detection results and the three-dimensional quantitative localization report, receive correction feedback from physicians, and use the corrected data as new training samples to incrementally learn the deep learning segmentation model.

Images