Recently, the team led by Zhou Benqing from the Department of Biomedical Engineering at the School of Engineering at Shantou University, together with Wang Zhongling's team from the First People's Hospital affiliated to Shanghai Jiao Tong University School of Medicine, published an article titled "AI-Assisted Tumor Boundary Delineation via Targeted Ultrasmall Iron Oxide Nanoprobe" in the internationally renowned journal Advanced Science (IF=14.1, CAS Zone 1 TOP). for High-Contrast HER2-Positive Tumor Imaging, with Shantou University as the primary institution.
Breast cancer is one of the most common malignant tumors in women, with HER2-positive breast cancer accounting for about 20%. It is characterized by strong aggression and a high risk of metastasis. Clinically, accurately identifying lesions and clearly delineating tumor boundaries is of great significance for diagnosis, determining surgical extent, and individualized treatment. MRI offers advantages such as non-invasive nature and high soft tissue resolution, but highly sensitive imaging usually relies on contrast agents. Traditional gadolinium-based T1 contrast agents have short circulation times, lack of tissue specificity, and potential safety risks. Ultra-small iron oxide nanoparticles (USIO NPs), due to their small particle size, good biocompatibility, and easily functionalized surfaces, can serve as potential T1-type MRI contrast agents. Meanwhile, the rapid development of deep learning artificial intelligence technology in medical image segmentation enables tumor regions to be automatically identified and reconstructed from 3D images. Therefore, a key question worth exploring is: Can we work synergistically from the two dimensions of "imaging probes" and "image deep learning" to enhance the true T1 signals of HER2-positive tumors while using artificial intelligence technology to precisely outline tumor boundaries, thereby achieving precise diagnosis and intraoperative navigation of breast cancer?
To address these issues, the study reported a trastuzumab (Tmab)-modified ultra-small iron oxide nanoprobe USIO@Tmab NPs, combined with the 3D nnU-Net deep learning framework for high-contrast MRI and boundary mapping of HER2-positive breast tumors. This system forms a collaborative strategy of "targeted imaging + microenvironment activation + AI segmentation": Tmab monoclonal antibodies recognize HER2 receptors, USIO nanoparticles are activated in the acidic tumor microenvironment and enhance T1 signals, and then 3D nnU-Net deep learning technology is used to automatically segment and visualize 3D MRI data.
Material characterization shows that the USIO core has a particle size of about 2.2 nm, maintains good dispersibility and superparamagnetic properties even after binding to Tmab antibodies, and has a protein loading efficiency of about 80.5%. This probe has excellent pH responsive MR imaging performance: at pH 7.4, its longitudinal relaxation rate r1 is only 1.43 mM-1·s-1; When pH drops to 5.0, r1 rises to 4.07 mM-1·s-1, an increase of about 2.85 times. The authors believe that acidic conditions promote USIO@Tmab structural dissociation, enhancing USIO interactions with surrounding water molecules, thereby shortening T1 and producing a brighter T1-weighted signal.
Cell experiments further validated the probe's targeting and safety. At a maximum iron concentration of 100 μg/mL, cell viability remained above 90%. In fluorescence imaging, USIO@Tmab can clearly bind SKBR3 cells with high HER2 expression, while signals are weaker in HER2-negative 4T1 cells. In vitro MRI also showed that the R1 value of 4T1-HER2 cells increased significantly over time and remained higher than that of HER2-negative 4T1 cells and normal MCF-10A cells, indicating that HER2-mediated uptake combined with the acidic tumor environment promotes signal activation.
In mouse experiments, after intravenous injection of USIO@Tmab NPs, the T1 signal in HER2-positive tumors gradually increased, peaking at about 240 minutes, with tumor R1 increasing by about 71.8% compared to before injection. Compared to commercial Gd-DTPA, USIO@Tmab has a more durable imaging window, with peak R1 about 25% higher. In bilateral tumor models, HER2-positive tumors have R1 at peak times about 23% higher than HER2-negative tumors, further demonstrating their in vivo targeted imaging capabilities.
For image analysis, the authors used 1,827 labeled MRI slices from 10 mice for training and evaluation using 3D nnU-Net technology, employing 50% cross-validation on an animal-based basis to avoid data leakage. The model obtained a Dice coefficient of 0.9336 and an IoU of 0.8762, with predicted boundaries highly consistent with manual annotation, and can also reconstruct three-dimensional tumor morphology based on continuous segmentation results.
More importantly, the tumor/normal tissue ratio in USIO@Tmab-enhanced images was about 2.72, which increased to 7.07 after 3D nnU-Net segmentation and image fusion, an increase of about 2.59 times. Here, AI does not alter the physical signals of the MRI itself, but instead further highlights tumor boundaries through automatic segmentation and mask fusion.
Overall, this work proposes a dual-layer enhancement strategy of "targeted/micro-environmental response contrast agent + 3D deep learning segmentation." The former enhances intratumoral MRI contrast through HER2 recognition and acidic microenvironment response, while the latter further highlights tumor margins and provides three-dimensional morphological information without altering underlying MR signals. Compared to relying solely on commercial contrast agents or only performing AI post-processing on existing images, this collaborative approach from material design to image computing is more conducive to improving the localization, boundary determination, and preoperative visualization capabilities of HER2-positive tumors. The core value of this study is not just "brightening MRI images," but establishing a complete chain across three levels: tumor-specific signal generation, automatic boundary recognition, and three-dimensional structural presentation, providing a technical framework worthy of further validation for precise imaging diagnosis, preoperative planning, and future surgical navigation for HER2-positive breast cancer.
This research was supported by projects such as the National Natural Science Foundation, Guangdong Provincial Natural Science Foundation, and the Li Ka Shing Foundation.
