Ultrasound Image Of Lymphoma Recognition Model And Interpretability Study

The lymphatic system is an important immune system that protects human health,and lymphoma is a malignant tumor caused by a lesion in the human lymphatic system.In recent years,the number of lymphoma patients in China has been increasing,which has seriously affected people’s health.Ultrasound-guided coarse needle aspiration biopsy is widely used because of its minimally invasive,convenient,and radiation-free characteristics,and it is more acceptable to patients because of its low price.Ultrasound images can show the lymphatic system more completely,which is helpful for doctors to analyze the type of lesion.In recent years,there has been an increase in research on automated lymphoma diagnosis and treatment systems,and most of these studies have focused on pathology images and electron computed tomography images.These images are of higher quality compared to ultrasound images,but they are also more expensive to obtain and are not suitable for early screening.The application to ultrasound images is basically a traditional statistical model.Statistical models require human processing,which is labor-intensive,while human extraction of features is more crude and the accuracy of diagnosis is not high.Moreover,these models generally lack interpretability and credibility.To solve the above problems.In this paper,we investigate the ultrasound image recognition model and interpretability of lymphoma,segment the images,and assist in classification according to the segmentation results,and design and implement a diagnosis and treatment system to assist doctors in diagnosing lymphoma species.The main research work of this paper has the following four aspects.1)Propose a semantic segmentation model based on self-attention mechanism and stable learningDue to the inherent acoustic characteristics of ultrasound images,the image quality is poor and there is much noise.According to the characteristics of ultrasound images,firstly,selfattention is added to the encoder network to alleviate the problem of information decay caused by multiple sampling,and then the weight of samples is calculated by a stable learning method,and the model loss function is optimized to remove the correlation between features,the model performance is effectively improved based on the baseline network Deep Labv3+.The model achieved 92.18% and 86.91% on the DICE coefficient and IOU score,respectively.2)Propose a classification model based on causal attention mechanism and feature fusionIn the ultrasound image of lymphoma,the background and the gray value of the lesion area are similar.The image of the lesion area is obtained through the above-mentioned semantic segmentation model,and the features of the two are extracted in parallel for fusion,so that the model pays more attention to the lesion area.The false correlation between features was removed by causal attention,and the classification accuracy of lymphoma was effectively improved based on the baseline network Res Net.The model achieved 90.4%,88%,and 89% in accuracy,recall,and F1-score,respectively.3)Use the counterfactual explanation method to explain the model prediction resultsA counterfactual explanation method is an instance-level explanation method that does not modify the model.According to the model and data characteristics of this paper,the pixel fragments are obtained according to the original image,and the original image of the segmentation model is disturbed.The classification model disturbs the original image and the lesion image at the same time,and the pixels in the disturbed area are set to 0,which is the model prediction.The results generate an instance-level explanation,showing the doctor which part of the image is critical to the model’s judgment.4)Design and implement an automatic recognition system for lymphoma ultrasound imagesThe system takes the model proposed in this paper as the core.By uploading the ultrasound image of the patient,the semantic segmentation result of the patient image is generated,the lesion area is obtained,and the category of lymphoma and the interpretation result are given to assist doctors in diagnosis.