United Imaging Intelligence Unveils Seven MICCAI-Selected Innovations Targeting Early Alzheimer's Detection and Clinical Efficiency

Pioneering a Dynamic Functional Network Model of Brain Regions: A Novel Perspective for the Precise Diagnosis of Mild Cognitive Impairment and the Prevention and Treatment of Alzheimer’s Disease, the “Memory Eraser”

Complete intelligent multi-region organ segmentation in as fast as 0.7 seconds with a single click, achieving a thousand-fold speed increase over manual delineation to “race against time” for patients.

Image fusion and registration completed in seconds, enabling precise tumor localization for needle biopsy with sub-millimeter accuracy.

Recently, the acceptance results for papers at the 2019 International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2019) were announced. Seven academic achievements from UNITED IMAGING stood out among nearly 2,000 submissions and were accepted by the conference. These achievements cover hot AI topics such as diagnosis and assessment of brain diseases, intelligent organ segmentation, image fusion and registration, and enhancement of system image resolution. Notably, an original algorithm developed by UNITED IMAGING for diagnosing early mild cognitive impairment—the Brain Region Dynamic Functional Network Model Algorithm—was selected for an oral presentation at the conference. All these achievements have been gradually implemented across China and are widely used in clinical practice and research projects at hospitals.

As a premier international conference on medical imaging, MICCAI (International Conference on Medical Image Computing and Computer Assisted Intervention) holds significant global influence and academic authority, long serving as the bellwether for the fields of Medical Image Computing (MIC) and Computer-Assisted Intervention (CAI). This year, fueled by the rapid advancement of AI, the conference attracted research teams from 134 top-tier universities and research institutions worldwide. The number of paper submissions increased by 70% year-on-year, reaching a new record high. Adhering to its consistently rigorous review standards, the conference maintained a final acceptance rate of only 31%.

Mild Cognitive Impairment (MCI) is considered a prodromal stage of Alzheimer’s disease (AD), and early identification of MCI is key to the prevention and treatment of AD. Patients with early-stage MCI do not exhibit significant structural brain changes, making diagnosis challenging and often leading to oversight.

UNITED IMAGING proposes a functional MRI processing technique based on graph convolutional networks to analyze dynamic brain region connectivity patterns for the diagnosis of early-stage mild cognitive impairment (MCI). Traditional MCI diagnostic methods primarily rely on cognitive assessment scales, such as the Mini-Mental State Examination, and are heavily dependent on empirical judgment and subjective evaluation, making it difficult to accurately assess structural brain changes or scientifically diagnose early-stage MCI. The UNITED IMAGING MCI computer-aided diagnosis system can rapidly detect subtle differences in the dynamic changes of brain region connectivity, automatically estimate patient gender and age, and precisely identify patients with early-stage MCI, thereby supporting the prevention and treatment of Alzheimer’s disease.

Osteoarthritis is a degenerative joint disease caused by various factors such as wear and tear, trauma, and joint deformity. In China, the prevalence of osteoarthritis exceeds 50% among individuals aged 60 and above, reaching as high as 68% in those aged 65 and older. Osteoarthritis can affect joints throughout the body, with the knee joint being the most commonly involved.

To enable more precise quantitative analysis of joint tissues and assist physicians in understanding the pathological features of knee osteoarthritis at different stages, UNITED IMAGING employs a Dice loss function based on a gradient adaptive mechanism for intelligent MRI-based knee segmentation. Manual organ segmentation by physicians is time-consuming, often taking tens of minutes to several hours, resulting in low work efficiency. In contrast, UNITED IMAGING’s intelligent MRI knee segmentation method can segment tissues of varying categories and difficulty levels within seconds, providing quantitative data to support clinical diagnosis and significantly improving both segmentation accuracy and diagnostic efficiency.

Pneumothorax is a pulmonary abnormality caused by air leakage into the space between the lung and the chest wall. Spontaneous pneumothorax is one of the respiratory emergencies in pulmonology. If patients with persistent or recurrent pneumothorax do not receive timely or appropriate diagnosis and treatment, their lung function may be impaired, potentially leading to shock and life-threatening complications. Therefore, for patients with pneumothorax, how can clinicians seize the “golden hour” and race against time to complete diagnosis and treatment?

Currently, chest X-rays are one of the most commonly used methods for diagnosing pneumothorax. However, as a routine and high-volume examination in hospitals, there is often a significant delay between image acquisition and the issuance of a diagnostic report by physicians. Furthermore, the diagnosis of pneumothorax and the formulation of treatment plans typically rely on visual interpretation of images, making it difficult to achieve precise quantification of the pneumothorax region and size. The UNITED IMAGING Pneumothorax Diagnosis and Segmentation Model is trained using image-level and limited pixel-level annotated data. It can rapidly and accurately segment the areas of pneumothorax and lung lobes, quantify the severity of pneumothorax based on the area ratio, and assist physicians in timely diagnosis and treatment of critically ill patients.

Mild cognitive impairment (MCI) can be categorized into two types based on the progression of the disease course: stable MCI, characterized by relatively stable cognitive function; and progressive MCI, which evolves into Alzheimer’s disease over time. The annual conversion rate from MCI to Alzheimer’s disease is 10–15%. Due to this high conversion rate and the irreversible nature of the condition, MCI is considered the optimal stage for patient intervention. Therefore, accurate differentiation between stable and progressive MCI is crucial in clinical practice.

Currently, most research efforts focus on extracting key features from medical images for precise diagnosis, while the intrinsic correlation between the task of localizing key regions in images and the task of extracting disease-specific features is often overlooked. Leveraging this correlation, UNITED IMAGING has developed a diagnostic method for cognitive impairment disorders that employs an iterative attention-focusing strategy derived from localization tasks. This approach enables rapid identification of affected brain regions and simultaneous classification of disease types, significantly improving both localization accuracy and diagnostic precision. It assists physicians in diagnosing mild cognitive impairment and assessing its progression, thereby facilitating timely intervention and treatment.

The human brain consists of two approximately symmetrical left and right hemispheres, with the boundary between them referred to as the midline. Under normal circumstances, the midline appears as an approximately straight line on axial brain images; however, certain neurological conditions, such as traumatic brain injury, can cause a shift in the midline. In such cases, the degree of midline shift serves as an important quantitative indicator for these conditions. Accurate assessment of this metric requires delineation of the midline, a process that is time-consuming when performed manually and highly dependent on the clinician’s expertise.

To improve segmentation efficiency and reduce the workload on physicians, UNITED IMAGING has developed a fully automated method for brain midline delineation based on regression and multi-scale feature fusion technologies. Traditional delineation methods mostly rely on symmetry priors between the left and right hemispheres to extract relevant features for constructing the brain midline; however, these methods struggle to accommodate cases involving severe midline shift and reduced bilateral symmetry caused by serious brain diseases. The fully automated brain midline delineation method from UNITED IMAGING utilizes a regression-based midline detection network, which can easily handle automatic midline delineation in cases of severe brain disease. Furthermore, it provides precise quantitative data on midline shift, offering an accurate methodological basis for the auxiliary diagnosis of related conditions.

The goal of thermal ablation surgery for liver tumors is to ablate tumor tissue as completely as possible while preserving surrounding healthy tissue. Needle insertion during the procedure is performed under image guidance, and accurate puncture requires registration of preoperative and intraoperative images. However, differences in imaging modalities between preoperative and intraoperative scans, along with significant deformation of tissues such as the liver, substantially increase the difficulty of image registration. Leveraging an MR-CT image registration method based on image synthesis and inpainting, UNITED IMAGING’s liver tumor fusion registration technology can rapidly and accurately fuse and align preoperative CT images with intraoperative MR images, completing tumor localization and image registration within 3 seconds. This assists physicians in performing thermal ablation of liver tumors with greater efficiency and precision.

In clinical practice, constrained by factors such as hardware limitations and scan time, magnetic resonance (MR) images are typically constructed by stacking multiple sparsely spaced two-dimensional acquisitions. This results in extremely low resolution in the direction perpendicular to the acquisition plane for the generated three-dimensional MR images. To some extent, this poses challenges for three-dimensional image processing and visualization in both clinical practice and scientific research. Leveraging sparse fidelity constraints and regularization methods, UNITED IMAGING employs deep learning to reduce the slice spacing of MR images without requiring actual high-resolution training data, thereby enhancing MR image resolution. The resulting images are comparable in quality to those obtained using genuine high-resolution training data.