Sichuan Cancer Hospital Licenses IVIM-Based Radiomics Prediction System for RMB 1 Million Plus Royalties

Recently, Sichuan Cancer Hospital released a public notice on the transformation of scientific and technological achievements, proposing to transfer its patented technologies through exclusive licensing.“A Tumor Radiomics Prediction System Based on IVIM Quantitative Parameter Analysis”The relevant patents are licensed to industry partners for use, with a licensing fee ofRMB 1 million + 2% of the actual annual sales revenue of patented products. The inventor of this patent isProfessor Lang Jinyi and his team。

This technology involves a tumor radiomics prediction system based on IVIM quantitative parameter analysis. The systematic approach enables rapid and more accurate estimation of IVIM parameters, faithfully reflecting intratumoral water molecule motion and tumor heterogeneity, while integrating radiomics for precise prediction.

Assessment of Tumor Heterogeneity and Precise Prediction of Subtypes,This remains a core challenge in clinical diagnosis and treatment. The intravoxel incoherent motion (IVIM) model, derived from diffusion-weighted imaging (DWI), was originally developed as a key tool for the non-invasive analysis of tumor microcirculation. However, existing IVIM parameter estimation techniques suffer from numerous limitations, which severely constrain their clinical utility.

Conventional IVIM parameter fitting methods (e.g., least squares, conventional neural networks) face core technical bottlenecks: On one hand, physics-informed neural networks (PINNs) rely on multilayer perceptrons to fit nonlinear functions, which are prone to catastrophic forgetting, leading to insufficient accuracy in voxel-level parameter estimation and failing to accurately reflect the heterogeneity of water molecule motion within tumors; on the other hand, the superposition of image noise interference and limitations of the fitting models substantially reduces the validity of the extracted radiomics features, making it difficult to support precise tumor analysis.

From the perspective of clinical application,Existing technologies still exhibit significant practical limitations. Clinicians must integrate multiple sets of IVIM parameter images to determine tumor molecular subtypes, a process that not only demands extensive diagnostic experience but is also susceptible to subjective interpretation and lacks standardized guidelines. Although DWI has become a routine sequence in tumor imaging, inaccurate parameter estimation and ineffective feature extraction prevent the full utilization of acquired data. This results in ambiguous assessments of tumor heterogeneity and low accuracy in subtype prediction (with traditional methods yielding an AUC of only 0.68), thereby failing to meet the requirements for formulating personalized treatment plans and monitoring therapeutic efficacy.

In addition,In multi-b-value DWI image processing,Conventional techniques suffer from suboptimal registration accuracy and time-consuming parameter calculations, which further reduce the efficiency of clinical applications. These issues collectively hinder the clinical translation of IVIM technology; therefore, there is an urgent need for an integrated solution that combines high precision with strong practicality to address the pain points across the entire workflow, from image analysis to disease prediction.

Precisely those present in tumor radiomics analysis“Inaccurate IVIM parameter estimation, difficulties in voxel-level analysis, and significant subjective interference”...and other clinical pain points have prompted Sichuan Cancer HospitalHuang Na, Lang JinyiThe team will carry out targeted technical breakthroughs. The core advantage of the patented technology “Tumor Radiomics Prediction System Based on IVIM Quantitative Parameter Analysis,” which is being commercialized in this instance, lies in"Improved Physics-Informed Neural Network + Full-Process Automated Analysis" to Build a Complete Solution,From parameter estimation to disease prediction, a full-chain technological innovation has been achieved, thoroughly breaking through the limitations of traditional IVIM technology, namely “poor fitting performance, ineffective feature extraction, and experience-dependent prediction.”

This technology was the first toIn terms of core modelsAchieved a disruptive breakthrough—innovatively developedPhysics-Informed Neural Networks Based on Improved mspline Functions, successfully overcoming the limitations inherent in traditional multilayer perceptrons“Catastrophic Forgetting”andPoor Performance of Nonlinear FittingThese two core challenges.

Traditional Physics-Informed Neural Networks (PINNs) rely on conventional activation functions, making it difficult to accurately capture the complex dynamics of tumor water molecule motion. In contrast, this technology introduces an improved mspline function—a fusion of the SiLU nonlinear activation function and B-spline linear combinations—along the “edges” of neuronal connections. This approach not only enhances the model’s ability to fit complex data but also prevents information loss during feature computation.

Meanwhile, the output layer employs a Sigmoid activation function to constrain the estimated ranges of IVIM quantitative parameters (pure diffusion coefficient Dt’, perfusion coefficient Dp’, and perfusion fraction Fp’), ensuring that the parameter values align with clinically relevant physiological significance and effectively reducing interference from image noise.

Experimental validation has demonstrated that the model’s parameter estimation can achieve an AUC value of0.81, significantly outperforming the conventional least squares method (AUC = 0.68), and enabling a more accurate reflection of the differences in water molecule motion within tumors as well as their heterogeneity characteristics.

InEnd-to-End Automated AnalysisField, this technologyEstablished a closed-loop system of “image acquisition—registration—processing—prediction”, significantly lowering the threshold for clinical operations and reducing interference from subjective factors.

First, optimize the image acquisition and registration processes.For DWI images with different b-values, using the b=0 image as a reference, precise registration is performed through optimization algorithms such as translation, rotation, and scaling to ensure accurate pixel-to-pixel matching across multiple image sets, thereby providing a robust data foundation for subsequent parameter estimation. Meanwhile, clear configuration standards for scanning parameters (e.g., 3T field strength, slice thickness ≤5mm, in-plane resolution ≤2.5mm) are established to guarantee high quality and consistency of the acquired images.

Second, innovate the radiomics feature processing workflow.By acquiring the lesion region, performing multi-dimensional feature extraction (shape, first-order, and texture features), concatenating features, and eliminating invalid features using the Recursive Feature Elimination (RFE) method, core features highly correlated with tumor subtypes are selected to avoid interference from redundant information on prediction results. Finally, the optimized features are input into an SVM prediction model to achieve objective classification of tumor subtypes in a data-driven manner, without relying on physicians' subjective judgments, thereby significantly improving the consistency and accuracy of predictions.

Furthermore, this technology inClinical Utilityon“High-Efficiency Adaptation + Broad Applicability”significant advantages. In terms of efficiency, the model employs Kaiming initialization and the ADAM optimizer to achieve rapid convergence, enabling swift processing of multi-b-value DWI images. Compared with traditional methods, it substantially reduces the time required for parameter estimation and analysis.

InApplicable ScenariosIn this regard, the system is compatible with 1.5T/3T magnetic resonance imaging (MRI) scanners, supports image analysis of tumors in multiple sites such as the breast, lung, and liver, and demonstrates exceptional performance in predicting subtypes such as triple-negative breast cancer.

Meanwhile, the system adopts a modular design with seamless signal interconnections between modules. Clinicians need only input DWI images to automatically complete the entire workflow—from registration and parameter calculation to disease prediction—without requiring complex operations, thereby meeting the clinical needs of laboratory and radiology departments across hospitals at all levels.

Currently, in response to the core needs in precision oncology diagnosis and treatment—namely, “low efficiency of imaging diagnosis, incomplete lesion detection, and poor coordination between diagnosis and treatment”—medical device manufacturers and artificial intelligence companies both domestically and internationally have established a diversified R&D landscape centered on multi-modal image analysis and AI-assisted diagnosis and treatment.

United Imaging HealthcareCentering on the R&D of full-chain imaging equipment, we have established a comprehensive layout in the field of intelligent tumor imaging analysis, integrating “hardware + software + large models.” In 2025, the company released“Yuanzhi” Medical Large Language Model. Trained on tens of millions of medical imaging datasets, this large imaging model supports over 10 imaging modalities and more than 300 image processing tasks, demonstrating high accuracy in tasks such as tumor lesion segmentation and diagnosis of complex lesions.

Developed Based on This Large Language Model"Imaging AI Agent", achieving the "single-scan, multi-condition assessment" capability for chest CT, automatically detecting 37 common thoracic diseases and abnormalities, while also supporting voice-enabled intelligent drafting of diagnostic reports, thereby fundamentally transforming the traditional image interpretation workflow.

In the context of oncology diagnosis and treatment,United Imaging and Sun Yat-sen University Cancer CenterJointly DevelopedAI Detection System for Brain and Bone Metastases, and has been deployed in over 400 hospitals across China. The relevant technologies have been integrated into the company’s high-end equipment series, such as 3.0T MR systems, completing clinical validation and achieving large-scale market release, thereby providing full-process support for precise tumor detection and treatment efficacy monitoring.

Shukun TechnologyFocusing on multimodal image fusion analysis for oncology, its AI-assisted diagnostic systems for tumors in the lungs, breasts, and other areas have received Class III medical device approval from the NMPA. These systems enable automatic lesion detection, benign-malignant differentiation, and risk stratification, and have been deployed in over 1,000 hospitals across China.

Deepwise MedicalDedicated to in-depth exploration of early cancer screening and precision diagnostics, it has launched AI-powered cancer screening solutions covering multiple sites, including the lungs, liver, and breast. By achieving compatibility with equipment manufacturers such as GE and Siemens, it enables efficient utilization of imaging devices across brands. Its related products have completed multi-center clinical validation and are currently in the stage of large-scale market promotion, further expanding the industrialization pathway for intelligent analysis of oncological imaging.

In summary, both industry and academia are currently striving to promote the upgrade of tumor radiomics technology from single-point auxiliary diagnosis to full-process diagnostic and therapeutic empowerment through the deep integration of “device hardware + intelligent algorithms,” thereby accelerating its clinical translation and implementation.