10a. In one embodiment, the labeled plurality of ultrasound images are manually labeled.

In step S42, the teacher model 10a is used to label the unlabeled plurality of ultrasound images to obtain the plurality of ultrasound images with pseudo labels.

In step S43, the plurality of ultrasound images with pseudo labels generated by the teacher model 10a are filtered to filter out pseudo labels with poor quality. In one embodiment, the teacher model 10a is provided with a threshold. If the confidence index of a label recognized by the teacher model 10a is lower than the threshold, said label will be automatically filtered out.

In step S44, the plurality of ultrasound images with pseudo labels having better quality are added to the labeled plurality of ultrasound images to obtain the combined plurality of ultrasound images, and the combined plurality of ultrasound images are used as a new training data set, thereby using the new training data set to train a student model 10b.

In step S45, if the performance of the student model 10b is higher than that of the teacher model 10a, then the student model 10b is used as a new teacher model 10a for optimization, and returns to step S41 until the training is over, thereby obtaining the trained teacher model 10a.

In another embodiment, the image recognition module 12 establishes and trains the object detection model 12a by using a Convolutional Neural Network (CNN) in deep learning according to the plurality of training images. In other words, the object detection model 12a of the present disclosure is a You Only Look Once (YOLO) object detection model in the CNN so as to perform deep learning training on the object detection model 12a based on the plurality of training images, such that the trained object detection model 12a can recognize the positions and names of the plurality of anatomical structures from the image to be recognized, wherein the object detection model 12a includes the following two aspects.

In the first aspect, the object detection model 12a may be an object detection model based on YOLOv4, and as shown in FIG. 5, the framework of the object detection model 12a of YOLOv4 includes a Backbone, a Neck and a Head.

Specifically, the Backbone is CSPDarkNet53, such that the neural network framework can obtain richer gradient fusion information and reduce the amount of computation, wherein as shown in FIG. 6, the specific process of CSPDarkNet53 includes the following steps.

In step S61, after CSPDarkNet53 inputs one of the plurality of training images, a feature map is obtained after the operations of convolution and pooling through the base layer, such that the feature map is used as the input of the neural network, and the feature map is divided into a first part and a second part.

In step S62, the second part of the feature map is convolved by a dense block and the number of channels is halved to obtain more feature information.

In step S63, the space size and the number of channels of the second part of the feature map are reduced by the first transition layer.

In step S64, after combining the first part of the feature map and the processed second part by the second transition layer, the space size and the number of channels of the reorganized feature map are also reduced to obtain a new feature map, wherein the new feature map contains low-level features and high-level features, which are conducive to improving accuracy of target detection.

In detail, through CSPDarkNet53, the object detection model 12a enables the present disclosure to achieve: (1) increasing the learning ability of the CNN and maintaining accuracy even if the model is lightweight; (2) removing the computing bottleneck structure with high computing power (reducing computing); and (3) reducing memory usage.

Moreover, the Neck is SPPnet+PANnet. Through the Neck, the receptive field is expanded and information of feature maps with different scales is fused (better feature fusion).

Specifically, as shown in FIG. 7A, the Spatial Pyramid Pooling (SPP) net is a pyramid network, which is connected to the spatial pyramid pooling layer after the last convolutional layer. The spatial pyramid pooling layer proposes a structure that maps and converts a feature with arbitrary size into feature vectors with fixed length, and then combines them into the following full connection layer whereby the input to a neural network may use images with arbitrary size to generate a fixed-size output. SPPNet extracts multi-layer features by spatial pyramid pooling to make target detection more accurate and faster and to improve the resilience of neural networks. Similarly, SPPNet combines global and sub-region information by maximum pooling layers with different sizes to extract feature vectors with different receptive fields. Accordingly, SPPNet has a better ability to describe detailed features and enhances accuracy of detecting different types of objects.

Further, in the backbone network, deep high-level features have more semantic meanings, while low-level information is more content descriptive. Therefore, using multi-layer features to make a discriminative/discerning pyramid network is very important for accuracy of object detection. As shown in FIG. 7B, Path Aggregation Network (PANet) improves the feature hierarchy based on Feature Pyramid Networks (FPN). The low-level information in deep layers is enhanced by an additional bottom-up path, and adaptive feature pooling is introduced to integrate features from all levels to improve predictive ability.

Finally, Head is used for final detection, which applies anchor boxes on the feature map and produces a final output vector with class probabilities, object scores and bounding/encompassing boxes. As shown in FIG. 7C, there are three output heads, for example, the strides are 8, 16 and 32 respectively, such that the large output feature map detects small objects, and small output feature map detects large objects.

In the second aspect, the object detection model 12a may be an object detection model based on YOLOv5, and as shown in FIG. 8, the framework of the object detection model 12a of YOLOv5 also includes a Backbone, a Neck and a Head. In addition, the parts of this aspect that are the same as those of the first aspect will not be repeated herein.

Specifically, as shown in FIG. 9A, the main difference between the second aspect and the first aspect is that the backbone of YOLOv5 in the second aspect further includes a Focus framework (also known as SpaceToDepth), and the Focus framework does not appear in YOLOv3 and YOLOv4. As shown in FIG. 9B, the Focus framework is used for slicing operations such that through the Focus framework, the original input image (e.g., the image to be recognized) is sliced into a feature map and then a convolution layer with 32 filters is applied, and final output is a feature map for the object detection model 12a to subsequently detect the plurality of anatomical structures in the image to be recognized. As such, the design of the Focus framework can reduce parameters of neural networks, reduce the size of the object detection model