| With the rapid pace development of modern society,the frequency of diseases is also increasing year by year.In order to carry out accurate disease diagnosis,it is necessary to take images of key lesion areas through corresponding medical imaging-assisted technology,for radiologists to observe and analyze and issue corresponding diagnostic reports.However,at present,regardless of the size of the hospital,there are generally phenomena such as insufficient doctor qualifications or manpower,lack of imaging equipment,and heavy tasks of image analysis and report issuance,resulting in a huge workload for doctors and even missed and misdiagnosed cases.In order to reduce the workload of doctors while ensuring the quality of medical diagnosis reports,researchers have introduced artificial intelligence methods such as machine learning and deep learning into the medical field,and used computer-aided diagnosis methods to realize the automatic generation of medical image diagnosis reports.Due to the low resolution of medical images and the blurred boundaries of internal tissues,the diagnostic reports generated by current medical image diagnostic report generation algorithms have problems such as low quality,low diversity and low authenticity.By using the data fitting ability of the Generative Adversarial Network,it simulates the doctor’s understanding of medical images and diseases,and further generates medical diagnosis reports with strong readability and high accuracy.Based on the Generative Adversarial Network,this paper proposes two improved automatic generation algorithms for medical image diagnostic reports to improve the quality of generated diagnostic reports.The main research contents of this paper are as follows:(1)The current research status of algorithms for natural image caption and medical image diagnostic report generation was investigated in detail,and the current excellent algorithms and mainstream evaluation metrics were summarized.Relevant algorithms were classified and analyzed based on network model categories and whether reinforcement learning was used.Finally,current challenges and future research directions were discussed.(2)The Cap GAN model is proposed based on conditional Generative Adversarial Network and attention mechanism,which can improve the quality of diagnostic report generation and maintain clinical language style and clinicopathological information.In order to improve the quality of diagnostic report generation,the conditional Generative Adversarial Network is combined with the attention mechanism to improve the model’s attention to important lesion areas and important report texts.At the same time,a language style evaluator is introduced to make the generated diagnosis report consistent with the language style of the original report.Simulation results show that the proposed algorithm has a significant improvement.The experimental results are evaluated by the BLEU metric.In the Open-i dataset,the BLEU-1 score increases by an average of 10.15%,and the BLEU-2 score increases by an average of 13.06%.Similarly,the average growth rates of the BLEU metric scores in the LGK dataset are 19.30%,29.14%,and 25.73%,respectively.In addition,the language style evaluation results have also been improved subjectively.(3)The XLA-GAN model is proposed based on the conditional Generative Adversarial Network and the linear attention mechanism,which uses the high-level feature information of medical images and diagnostic reports to obtain fine-grained image and semantic feature representations to further improve the quality of diagnostic report generation.The linear attention module is introduced into the image feature encoding and diagnostic report decoding modules of the generator,and the bilinear pooling operation is used to obtain highorder fusion features,and the joint representation of image features and report generation hidden states is further effectively inferred.At the same time,a diagnostic report discriminator based on a Recurrent Neural Network is designed,and the reinforcement learning mechanism is used to optimize the alternate training of the Generative Adversarial Network,thereby improving the accuracy and readability of the generated medical diagnostic report.The simulation results show that the proposed algorithm achieves an average growth rate of about 8% to 30% in most of the evaluation metrics scores in the two datasets.In addition,the analysis of subjective experimental results shows that the XLA-GAN model has a certain improvement in accuracy and readability,and is more in line with the language style of clinical medical reports. |
