3.2. Core Patents & Key Technologies of Info Mining

1) 3D GAN Technology (antagonism generation) for Image Data

In the training of DL-deep learning, due to the lack of data, it is necessary to expand the data to assist the model training. At present, GAN, the existing data expansion technology, only can generate 2D spaces. If it is used to train models, it will result in: (1) only the 2D method can be used to conduct learning, and 3D DL-deep learning method cannot be used. (2) the generated 2D image will present an inauthentic structure and texture in 3-dimensional space. 3D GAN technology, which is suitable for image data, will generate 3D objects. Only the objects generated in 3D spaces can be maximally similar to the real 3D objects (such as lung and brain). In training, it can effectively reduce errors.

2) Dl-deep Learning of 3D Signals for Image Data

The conventional training platform of DL-deep learning is suitable for 2D signals, and for 3D signals training, its speed is very low due to the large memory consumption. Now, in combination with the particularity of image data, we adopt sampling technology that is especially suitable for 3D space to obtain a faster convergence rate under the same sample size, or to reduce the sample size under the same rate requirement. Various data demonstrate that the performance of 3D DL-deep learning is superior to that of 2D DL-deep learning.

3) Image Format Analysis Technology of Medical Cases and Medical Bills Based on Semi-supervised Learning

The format of medical cases and medical billing differs in thousands of ways, so we can't simply adopt the traditional template method. Therefore, we proposed the feature extraction method based on semi- supervised learning to analyze the image features and judge the bills' image formats. The image format feature is extracted by the non- supervised learning and the gradient space characteristics combination, and the method adopts the extreme learning machine approach to realize the division and determination of the image content areas.

4) Interpretation Technology of Words Meaning of Laboratory Sheet Based on Recurrent Neural Network

There often exists ambiguity and polysemy in the medical vocabulary, which increases the difficulty of recognizing and understanding medical terminology. We propose the language model based on the recurrent neural network, and by the construction of the probability model between words, we can realize the understanding and comprehensive analysis of medical terms. The technology also has the feature of fast self-learning, which can iterate in applications to improve the understanding ability of the system.

5) Medical Vocabulary Recognition Technology Based on Convolution Neural Network

There is distortion and quality degradation in the image of medical bills acquired by taking photos, and the existing character recognition core functions can't effectively meet the recognition accuracy requirement. Through the training of the sample dataset of labeled entry and image blocks, we construct a multi-block

identification model based on the convolution neural network model, which shows good robustness to medical vocabulary and image blocks with degraded and noise pollution. By building a richer sample set, we will further improve the robustness of medical vocabulary recognition technology.