Sun Yat-sen Memorial Hospital Seeks Transfer of Bladder Cancer Lymph Node Metastasis AI Diagnostic Patent for RMB 500,000

Recently, Sun Yat-sen Memorial Hospital of Sun Yat-sen University released a public notice on patent transfer, proposing to transfer its jointly held"Pathological Image Recognition Method, Device, and Medium for Lymph Node Metastasis of Bladder Cancer"The 50% equity interest in the patent is transferred through a listed transaction, with the transaction amount being500,000 yuan. By integrating non-uniform resampling techniques with a target segmentation network, this patent can automatically identify suspected cancer cell regions in pathological images of bladder cancer lymph node metastases. It effectively addresses the pain points of traditional manual diagnosis, such as being time-consuming and labor-intensive, as well as the high rate of missed micro-metastases. This innovation significantly improves the efficiency and accuracy of pathological image recognition, providing key technical support for tumor staging, treatment guidance, and prognosis assessment in patients with bladder cancer.

Bladder cancer is one of the most common malignant tumors in the clinical urinary system,Tumor Lymph Node Metastasis StagingIt serves as the core basis for determining disease progression and formulating treatment plans. When cancer cells spread to lymph node tissue via the lymphatic circulation, they form metastatic foci of varying sizes. Among these, “micrometastasis” specifically refers to the presence of a small number of cancer cells scattered within the lymph nodes. The detection of such lesions directly influences the selection of subsequent treatment regimens and the assessment of patient prognosis.

For a long time, the diagnosis of lymph node metastasis in bladder cancer has relied heavily on the manual work of pathologists. Physicians must examine lymph node pathology slides slice by slice and region by region under a microscope, a process that not only consumes hours or even days but also carries a high risk of missed diagnoses. This is particularly true for micrometastases; due to the small number, scattered distribution, and minute size of cancer cells, the human visual system has extremely low sensitivity to such tiny targets. Consequently, missed diagnoses are highly likely due to visual fatigue or oversight, which can delay optimal treatment timing for patients and even compromise the precision of subsequent therapeutic strategies.

Currently, although certain image recognition technologies for computer-aided diagnosis have been applied in clinical practice, they still exhibit significant limitations. Conventional resampling methods typically employ a fixed-threshold approach to crop and magnify suspicious regions. This process often results in the loss of contextual information from surrounding normal cells and tissues, thereby hindering comprehensive assessment of lesion characteristics based on contextual cues. Furthermore, such techniques struggle to simultaneously magnify multiple scattered regions suspected of containing cancerous cells, often necessitating the output of multiple isolated regions of interest. This not only increases the interpretive workload for physicians but also reduces diagnostic efficiency and accuracy. In addition, traditional AI segmentation models lack targeted data augmentation during training, leading to insufficient generalization capabilities when faced with complex pathological images, thus failing to meet the urgent clinical demand for precise diagnosis.

The core innovation of this invention patent lies inBreaking through the limitations of traditional image recognition technology, we have constructed an efficient and accurate pathological image recognition system centered on “non-uniform resampling technology,” combined with AI segmentation networks and multi-dimensional data fusion strategies.Unlike conventional resampling methods that merely crop and magnify suspicious regions, this patent’s non-uniform resampling technology employs coordinate transformation functions and a weight calculation mechanism to automatically enlarge suspected lesion areas while fully preserving information on normal cells and tissues surrounding the suspicious regions. This provides critical support for physicians to assess lesion characteristics in context, thereby technically reducing the risk of missed diagnoses of micrometastases.

Another major innovation of this patent is the realization of“Synchronous Optimization of Multiple Scattered Suspected Regions”. In view of the characteristic that cancer cells may be dispersed within lymph node tissue, this technique does not output multiple isolated suspicious regions separately. Instead, it synchronously magnifies and integrates all suspected lesions into a single image patch through non-uniform resampling. This approach not only reduces the interpretation workload for physicians but also avoids misdiagnosis caused by analyzing isolated regions, thereby significantly improving diagnostic efficiency. Furthermore, the resampled image patches maintain the same resolution as those input to the target segmentation network, effectively reducing computational resource consumption while balancing processing speed and recognition accuracy.

At the model training level, patent innovations adopt“Dual-Dataset Fusion Training Strategy”By applying non-uniform resampling to the training image set and segmentation annotation maps, we generated a resampled image set and corresponding resampled segmentation annotations. These were combined with the original data for end-to-end segmentation training, incorporating a binary cross-entropy loss function to simultaneously optimize segmentation losses for both original and resampled data. This approach significantly enhanced the generalization capability of the target segmentation network (HRNet_w18) and its sensitivity in detecting small lesions. The final output, a segmentation confidence heatmap, intuitively visualizes pixel-level probability distributions of cancer cells, providing physicians with precise diagnostic references and achieving a triple breakthrough in “automated screening, precise localization, and efficient interpretation.”

Bladder cancer, as the fourth most common malignant tumor among men worldwide, itsEarly Diagnosis, Metastasis Detection, Recurrence Monitoring, and Pathological GradingIt directly impacts patients’ treatment plans and prognostic assessment. Traditional diagnosis and treatment rely on pathologists’ manual slide review or invasive examinations, which are time-consuming and labor-intensive, carry a high risk of missed diagnoses, and result in low patient compliance. In recent years, artificial intelligence (AI) has emerged as a key force in addressing this challenge, with multiple AI patents focused on bladder cancer diagnosis and treatment being introduced. These technologies cover core scenarios such as lymph node metastasis detection, recurrence monitoring, and pathological grading.

Northeast Forestry UniversityR&DIdentification Technology for Mediastinal Lymph Node Metastasis Based on a Dual-Control Routing Mixture of Experts Model, the core task is decomposed into two auxiliary subtasks: mediastinal region detection and lymph node metastasis prediction, thereby constructing a multi-task learning system. This technology employs a dual-control routing gating mechanism, wherein the feature routing branch extracts image-similar features (color, brightness, and frequency domain), while the task routing branch extracts task-specific features; these are then weighted and fused to achieve a comprehensive feature representation. Furthermore, a dual-dimensional gradient balancing algorithm is designed to address gradient conflicts and task dominance issues by aligning gradient directions and dynamically balancing gradient magnitudes. Although primarily applied to the identification of mediastinal metastases in lung cancer, its approach to multi-task collaboration and refined feature extraction provides significant reference for the identification of lymph node metastases in bladder cancer, establishing a technical paradigm for AI-based diagnosis of lymph node metastases across different cancer types.

Guangzhou Zhihui Matrix Life Technology Co., Ltd.launchedMultimodal Bladder Cancer Recognition System, integrating three MRI modalities—T2-weighted imaging, diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps—the system projects each image type into a high-dimensional feature space to generate embedding vectors. Global features are then extracted using a Vision Transformer (ViT) model, after which a fusion module adjusts dimensions and concatenates the features to produce fused feature vectors. Finally, a classification module performs dual classification based on both single-modality and multi-modal fused features, outputting predictions for bladder cancer recurrence. By fully leveraging complementary information across modalities—where T2-weighted images provide structural details, DWI reflects cellular density, and ADC quantifies the degree of diffusion—the system effectively distinguishes tumor extension from benign lesions. This non-invasive approach partially replaces invasive cystoscopy, reducing patient burden and improving follow-up compliance. The fusion-based classification accuracy of the fourth classification unit reaches 79%, enhancing the robustness and reliability of identification.

Kunming University of Science and Technology & Yuxi Normal UniversityJointly developedGrading Method for Multi-Sequence MRI Images of Bladder Cancer Based on the ResNet ModelFor MRI data from the T2WI and DCE sequences, tumor-containing images were selected based on regions of interest (ROIs) delineated by physicians. The lesion areas were preserved through central cropping, and data augmentation was performed using random rotation and flipping. Transfer learning was introduced based on the ResNet50 model, incorporating an MF multi-scale feature extraction module. Furthermore, a scSE attention mechanism was applied after multi-sequence feature fusion to enhance the learning of key features through channel and spatial attention. This technique addresses the limitations of relying on postoperative pathological biopsy for bladder cancer grading, enabling precise preoperative differentiation between high-grade urothelial carcinoma (HGUC) and low-grade urothelial carcinoma (LGUC), thereby providing critical reference for surgical strategy formulation and prognosis prediction.

In the future, AI-based diagnostic and therapeutic technologies for bladder cancer will move toward“Full-Process Coverage” “Deep Multimodal Integration” “Low Data Dependency”Through cross-scenario technological synergy, this evolutionary direction further reduces rates of missed and misdiagnoses while enhancing diagnostic and therapeutic efficiency. It provides patients with intelligent, full-cycle auxiliary support spanning screening, diagnosis, treatment, and follow-up, thereby ushering bladder cancer care into a new era characterized by precision, high efficiency, and minimal invasiveness.