1 Algorithm Introduction
With the development of artificial intelligence, computer vision technology has become increasingly widely used. For example, in video surveillance for face recognition and identity analysis, medical diagnosis for analyzing and recognizing various medical images, fine-grained visual classification, facial image attribute recognition, fingerprint recognition, scene recognition, etc. Computer vision technology is gradually penetrating into people’s daily lives and applications. However, traditional computer vision methods that manually extract image features and then perform machine learning are becoming increasingly inadequate for these applications.
Since 2006, deep learning has entered people’s sights, especially after AlexNet won the championship in the ImageNet Large Scale Visual Recognition Challenge in 2012. Deep learning has achieved remarkable development in the field of artificial intelligence, achieving great success in many fields such as computer vision, speech recognition, natural language processing, and multimedia. The biggest difference between deep learning and traditional pattern recognition methods is that it automatically learns features from big data rather than using manually designed features, and good features can greatly improve the performance of pattern recognition systems. In the field of computer vision, deep convolutional neural networks have become a research hotspot, playing a crucial role in computer vision tasks such as image classification, object detection, and image segmentation.
Deep Convolutional Neural Networks (DCNN) are a very powerful tool in deep learning, widely used in computer vision and image processing. Through multiple layers of convolution and pooling operations, it can effectively extract features from images and has strong expressive and generalization capabilities. DCNN is a deep learning model composed of multiple convolutional layers, pooling layers, and fully connected layers. Its core idea is to extract features from images through convolution and pooling operations and perform classification or regression tasks through fully connected layers.
2 Algorithm Principles
In DCNN, the convolutional layer is responsible for extracting local features of the image. It performs convolution operations on the input image by defining a set of convolution kernels (or filters), resulting in a set of feature maps. Each feature map corresponds to a convolution kernel, indicating the degree of response of that convolution kernel to the input image. Convolution operations can effectively capture local features such as edges and textures in the image. The pooling layer is used to reduce the size of the feature maps while retaining the main feature information.
To achieve better performance, the number of layers in convolutional neural networks is increasing. The deeper the deep convolutional neural network, the more parameters need to be learned, and the harder it is to optimize the network. Without good optimization methods, overfitting or underfitting issues may arise.
Overfitting means that the model has poor generalization ability; it fits well on the training set but poorly on the validation set. Simply put, the model learns too strongly from the training data and can recognize images in the training set well, but fails to recognize images outside the training set. There are two reasons for this: one is that the training set data is too small, and the other is that the number of training iterations is too high. There are mainly four methods to mitigate overfitting, as follows:
(1) Early stopping. After each iteration (epoch), calculate the validation error. If the error no longer decreases, stop training. This is a timely loss prevention method; the model’s generalization ability no longer improves, and continuing training would be a waste of time. However, relying solely on the error rate after one iteration is unscientific, as the error rate may rise or fall after this iteration. Therefore, the error rates after 10, 20, or more iterations can be used to determine whether to stop training.
(2) Data augmentation. This is the most direct and effective method to alleviate overfitting. Without high-quality, abundant data, it is impossible to train a good model. Data augmentation can be approached from two aspects: increasing data from the source, such as directly adding images to the training set during image classification, but this method is difficult to implement as it is hard to determine how much data to add; or modifying the original data to obtain more data, such as rotating the original images, adding noise to the original data, or cropping parts of the original data.
(3) Regularization. Regularization includes L0 regularization, L1 regularization, and L2 regularization, with L2 regularization being the most commonly used in machine learning. The L2 regularization term serves to reduce parameters, meaning a lower model complexity, which improves the model’s generalization ability.
(4) Dropout. Dropout, as a model ensemble method, randomly disables neurons with a certain probability, effectively reducing test error. Given an input, the network samples different structures that share a set of parameters. Since a neuron does not rely on specific neurons, dropout reduces complex co-adaptation between neurons and enhances the robustness of the network.
When the model fits poorly on the training set but well on the validation set, it is called underfitting. The reason for underfitting is that the model has not learned enough from the training data, and feature learning is insufficient, resulting in poor representation capability. Underfitting can be mitigated by the following methods:
(1) Adding other feature terms. Insufficient feature terms can lead to underfitting, which can be effectively resolved by adding feature terms. Methods for increasing feature terms include combinations, generalizations, correlations, etc., which are applicable in many scenarios.
(2) Adding polynomial features. For example, adding quadratic or cubic terms in a linear model enhances the model’s generalization capability.
3 Algorithm Applications
Deep convolutional neural networks have been widely applied in fields such as natural language processing and image recognition, achieving significant results. Similarly, in the field of traditional Chinese medicine, deep convolutional neural networks play an indispensable role. For example, in the study of TN staging for thyroid cancer, researchers developed a deep convolutional neural network model to improve the accuracy of preoperative TN staging assessment for thyroid cancer patients. They collected clinical data, pathological results, serum markers, etc., downloaded preoperative ultrasound images and results of thyroid nodules and cervical lymph nodes, and listed postoperative pathological TN staging and preoperative ultrasound TN staging. The ultrasound images of the patients’ thyroid nodules were preprocessed and used to construct a DCNN model for training, validating the model’s accuracy in assessing TN staging.
Research results show that the DCNN model achieved high accuracy in the validation set for T1 stage. Currently, the DCNN model can enhance the diagnostic accuracy of preoperative ultrasound in T1 stage, while more data training is still needed to achieve higher accuracy in other T stages; the accuracy of DCNN predicted N0 exceeds that of ultrasound diagnosis in predicting the absence of lymph node metastasis, guiding clinicians to perform more conservative preventive central lymph node dissection. The DCNN also shows some effectiveness in predicting N1b stage, potentially avoiding FNA examinations and providing better treatment for patients.
4 Summary
Deep Convolutional Neural Networks (DCNN) have achieved far better results in feature recognition-related tasks compared to traditional methods. Therefore, DCNNs are commonly used in image recognition, speech recognition, etc. However, due to the large structure of deep convolutional neural networks, which generally contain dozens of neural layers, each with hundreds to thousands of neurons, and the intricate interactions between neurons in any two layers, these two main factors make DCNNs difficult to understand and analyze. As a result, users find it challenging to learn how to design a good convolutional neural network from examples of failures or successes. Consequently, designing an effective neural network often requires a lot of trial and error.
[1] Wang Hui. Research on the Application of Ultrasound and Deep Convolutional Neural Networks in TN Staging of Thyroid Cancer [D]. Zhejiang Chinese Medical University, 2024. DOI:10.27465/d.cnki.gzzyc.2023.000058.
[2] “Powerful Tools in Deep Learning – Deep Convolutional Neural Networks (DCNN) Principles – CSDN Blog”. Retrieved on February 21, 2024.
[3] “Deep Learning (Deep Learning) – Deep Convolutional Neural Networks (Deep Convolutional Neural Networks) –
[4] “Introduction to Deep Learning – Overview of Deep Convolutional Neural Network Model (Deep Convolution Neural Network, DCNN) – CSDN Blog”. Retrieved on February 21, 2024.
Factor Analysis – To “Strike”, First Reduce Dimensions
Independent Component Analysis (ICA) – “Butchering” Mixed Data Signals
Graph Convolutional Networks – More Sophisticated Feature Extractors for Graph Data
Ancient and Modern Medical Case Cloud Platform
Providing over 500,000 ancient and modern medical case retrieval services
Supports manual, voice, OCR, and batch structured input of medical cases
Designed nine analytical modules, close to clinical practical needs
Supports collaborative analysis of massive medical cases and personal cases on the platform
EDC Traditional Chinese Medicine Research Case Collection System
Supports multi-center, online random grouping, data entry
SDV, audit trails, SMS reminders, data statistics
Analysis and other functions
Supports customized form design
Users can log in at: https://www.yiankb.com/edc
Free Trial!
Institute of Traditional Chinese Medicine Information Research, Chinese Academy of Traditional Chinese Medicine
Smart R&D Center for Traditional Chinese Medicine Health
Big Data R&D Department
Phone: 010-64089619
13522583261
QQ: 2778196938
https://www.yiankb.com
