WiMi Developed a Generalized Learning Algorithm of Auto-AIX

BEIJING, Aug. 15, 2023 /PRNewswire/ — WiMi Hologram Cloud Inc. (NASDAQ: WIMI) (“WiMi” or the “Company”), a leading global Hologram Augmented Reality (“AR”) Technology provider, today announced that a generalized learning algorithm of X-ray image analysis is developed for X-envisioned images, naming it Automatic Artificial Intelligence X-ray Image Analysis (Auto-AIX).

X-ray image analysis is a complex process involving the detection of various features such as bone density, organ shape and tissue density. Traditionally, this process has been performed manually by medical professionals who use their expertise to identify and analyze features. However, this method is time-consuming and can be subject to human error, leading to misdiagnosis and poor patient prognosis.

WiMi has been exploring the use of artificial intelligent algorithms to automate the process of X-ray image analysis. These algorithms are designed to learn from large datasets of X-ray images and can recognize patterns and features that are difficult or impossible for human experts to detect. By automating the X-ray image analysis process, AI algorithms have the potential to increase the speed and accuracy of diagnosis while reducing the workload of healthcare professionals. However, developing effective AI algorithms for X-ray image analysis requires a large and diverse dataset of X-ray images for training and validation. This dataset must be carefully selected and annotated to ensure that the AI algorithm can accurately recognize image features.

WiMi has developed a generalized learning algorithm for X-ray image analysis that is designed to learn from a diverse set of X-ray images to make it suitable for use in real-world applications. The algorithm is based on a deep neural network architecture that is trained using a large and diverse dataset of X-ray images. The datasets are collated and annotated to ensure that the algorithm accurately identifies features of interest such as bone density, organ shape and tissue density. In order to improve the generalization ability of the algorithm, several techniques such as data expansion and domain randomization are implemented. Data expansion consists of applying a series of transformations, such as rotation, scaling, and flipping, to the original X-ray images to create a larger and more diverse training dataset. Domain randomization involves adding random noise and perturbations to the training data, which helps the algorithm generalize to new and unseen X-ray images. The algorithm is designed to run on a range of hardware platforms, from conventional CPUs to high-performance GPUs. This makes it suitable for deployment in real-world environments where hardware resources may be limited or variable.

Auto-AIX includes data acquisition, generation, and annotation with generalized learning algorithms. Data acquisition, generation and annotation are key to building deep learning models. In the field of medical imaging, the collection and use of real data face many restrictions due to patient privacy and confidentiality. And Auto-AIX circumvents these restrictions by using computer-generated synthetic data. Specifically, it uses CT to model X-ray images, which gives the synthetic data a realistic appearance and detail, thus improving the accuracy of the model.

In Auto-AIX, the first step in generating synthetic data is to create a medical model, which can be modeled using a CT scan or a surgery tool. Then, by injecting noise and variations into the medical model, multiple samples can be generated that cover a wide range of situations and variations that may appear in real data. Finally, these samples are annotated, i.e., by manually labeling them with features and diseases. These annotations can be automatically applied to all other synthetic data, thus saving significant time and labor costs. This process is called “domain extension” in Auto-AIX, as it allows the synthetic data domain to be extended to a wider range of datasets.

Auto-AIX uses a generalized learning-based algorithm to build deep learning models. The advantage of this algorithm is that it can be trained using a large amount of synthetic data without the need for large amounts of real data. This means that Auto-AIX can train high-performance deep learning models even when there are difficulties and limitations in collecting real data.

Specifically, Auto-AIX uses domain randomization techniques to build algorithms based on generalized learning. The core idea of this technique is to improve the generalization ability of the model by introducing randomness in the appearance and features of the synthetic data. This randomness can be arbitrary, e.g., adding noise, perturbation, occlusion, etc. to the synthetic data. In this way, Auto-AIX can construct deep learning models with high generalization performance.

To evaluate the performance of Auto-AIX, the researchers conducted a series of experiments to compare the differences between deep learning models trained with Auto-AIX synthetic data and models trained with real data, and test the effect of using different synthetic data.

The experimental results of training AI models using SyntheX synthetic data show that the method can achieve performance comparable to, and in some cases even exceeding, real data training. Next, the trained AI model needs to be applied to real clinical X-ray image data for evaluation and deployment. Before applying the AI model to real data, the real data needs to be pre-processed to have a similar distribution to the synthesized data. This pre-processing method is called domain adaptation or domain transfer. The goal of the domain transfer method is to transfer the model from a source domain (synthetic data) to a target domain (real data) in such a way that the model performs optimally on the target domain. The main idea of domain adaptation is to learn a model that can be generalized over the target domain by modeling the difference in distributions between the source and target domains.

To apply AI models to real data, WiMi uses a domain adaptation method called Adversarial Discriminative Domain Adaptation (ADDA), which consists of two phases: the first phase is to train a source domain classifier and a target domain classifier to distinguish between source and target domains; the second phase is to train a domain adapter to transfer features from the source domain to the target domain so as to optimize the performance of the model on the target domain. classifier and target domain classifier to distinguish the difference between the source and target domains; the second stage is to train a domain adapter that transfers features from the source domain to the target domain so that the model performs optimally on the target domain. The ADDA method was used to transfer the model from the synthetic data domain to the real data domain. After domain adaptation, the performance of the AI model on real data is comparable to its performance on synthetic data, which shows that the domain adaptation method is effective.

WiMi’s AI model was applied to virtual clinical X-ray images and evaluated. The results show that the AI model can accurately identify a variety of diseases and abnormalities, including pneumonia, pulmonary nodules, and pulmonary effusions. In addition, the AI model can perform quantitative measurements, such as lung volume and nodule size. Overall, training the AI model using Auto-AIX synthetic data and domain adaptation using the ADDA method can greatly accelerate research and applications in the field of X-ray image analysis, bringing more opportunities and challenges to the healthcare field.

About WIMI Hologram Cloud

WIMI Hologram Cloud, Inc. (NASDAQ:WIMI) is a holographic cloud comprehensive technical solution provider that focuses on professional areas including holographic AR automotive HUD software, 3D holographic pulse LiDAR, head-mounted light field holographic equipment, holographic semiconductor, holographic cloud software, holographic car navigation and others. Its services and holographic AR technologies include holographic AR automotive application, 3D holographic pulse LiDAR technology, holographic vision semiconductor technology, holographic software development, holographic AR advertising technology, holographic AR entertainment technology, holographic ARSDK payment, interactive holographic communication and other holographic AR technologies.

Safe Harbor Statements

This press release contains “forward-looking statements” within the Private Securities Litigation Reform Act of 1995. These forward-looking statements can be identified by terminology such as “will,” “expects,” “anticipates,” “future,” “intends,” “plans,” “believes,” “estimates,” and similar statements. Statements that are not historical facts, including statements about the Company’s beliefs and expectations, are forward-looking statements. Among other things, the business outlook and quotations from management in this press release and the Company’s strategic and operational plans contain forward−looking statements. The Company may also make written or oral forward−looking statements in its periodic reports to the US Securities and Exchange Commission (“SEC”) on Forms 20−F and 6−K, in its annual report to shareholders, in press releases, and other written materials, and in oral statements made by its officers, directors or employees to third parties. Forward-looking statements involve inherent risks and uncertainties. Several factors could cause actual results to differ materially from those contained in any forward−looking statement, including but not limited to the following: the Company’s goals and strategies; the Company’s future business development, financial condition, and results of operations; the expected growth of the AR holographic industry; and the Company’s expectations regarding demand for and market acceptance of its products and services.

Further information regarding these and other risks is included in the Company’s annual report on Form 20-F and the current report on Form 6-K and other documents filed with the SEC. All information provided in this press release is as of the date of this press release. The Company does not undertake any obligation to update any forward-looking statement except as required under applicable laws.