Research On Medical Image Data Augmentation Method Based On Generative Adversarial Network

Small-scale datasets and limited labeled samples are the main challenges in the medical imaging domain.However,training a successful deep learning algorithm requires sufficient labeled data as support,but medical image data is difficult to obtain and label annotation requires expensive labor costs,which greatly limits its application in the medical domain.In medical imaging tasks,senior radiologists will make lesion labels based on their professional domain knowledge,but most annotations of medical images are time-consuming.When data is scarce,neural networks are prone to over-fitting problems,which is especially obvious on small-scale datasets.Traditional image data augmentation methods have a certain effect on alleviating the over-fitting problem,but compared with the original data,the performance improvement brought by synthetic data is still very limited.As the GAN model was proposed,researchers began to turn their attention to using GANs series models to generate images similar to the original dataset for data augmentation.Although a well-designed GAN variant model can generate images with the same quality as the original data,it still cannot avoid its inherent problems such as the inability to generate designated label images and the noise data contained in the synthesized images.In response to these problems,this paper conducts a series of innovative research on data augmentation methods based on the thyroid ultrasound image dataset.The main research results are as follows:1.Existing research is almost entirely based on learning model parameters and extracting deep features in a data-driven way,which may make the accuracy of the model much lower than human performance under real conditions such as small training samples.In response to this situation,this method designs a generative adversarial network model KACGAN,which integrates the expert experience and knowledge that radiologists have summarized for a long time into the model,deeply excavates and analyzes medical text information,integrates images and clinical data,and multi-dimensional comprehensive analysis to generate high-quality ultrasound image data.It is worth noting that this paper extracts expert domain knowledge from the standardized terms of imaging reports instead of extracting from ultrasound images,and designs a hetero-modal similarity loss to ensure the authority of domain knowledge features and the stability of model training.In this paper,the performance of the data augmentation method is verified in the task of thyroid lesion classification.The proposed KACGAN model has an accuracy of 91.46%,a sensitivity of 90.63%,a specificity of 92.65%,and an AUC of95.32%,which is better than the current classification of thyroid nodules.2.Existing research can generate high visual quality synthetic images based on GANs,but seldom consider how to ensure the characteristic quality of synthetic images or control the augmentation steps after image synthesis.Synthetic images without quality assurance may adversely change the data distribution and reduce model performance.In response to this situation,this paper proposes a feature-based selective data filtering framework,which filters out noise samples to prevent them from affecting the classification and generalization capabilities of the KACGAN model.In this paper,the original dataset is put into the Res Net18 model for training to obtain a feature extraction model,and then feature extraction is performed on the generated dataset to filter synthetic images that can be determined to be benign and malignant,and secondly,synthetic images that is sufficiently close to the centroid of the real images in the feature space is selected,remove noise data as much as possible.After setting up an experimental control group for verification,the results prove that the selective data filtering framework can further improve the quality of the dataset and the performance of the model classification.