Detection And Metastasis Prediction Of Abdominal Lymph Nodes Based On Deep Neural Network

Treatment of malignant tumours is based on the patient’s type of malignancy and the stage of the malignancy,while lymph node metastasis is an essential part of determining the stage of the malignancy.The gold standard for detecting lymph node metastasis is biopsy,which means additional invasive harm to the patient.Reading CT images manually can prevent additional patient harm,however,this work is inefficient and highly dependent on the physician’s professional experience.Therefore,exploring the use of deep learning methods to achieve lymph node metastasis prediction based on CT images has become one of the significant research directions in the intersection of medicine.Currently,there are still some problems and challenges in this study.On the one hand,there are problems with the data.First,the tiny volume of lymph nodes,the complex environment of the abdominal cavity,and the adhesion of lymph nodes to surrounding tissues make it difficult to detect lymph nodes.Secondly,the image resolution of lymph node metastasis dataset is normally much smaller than the input resolution of traditional network models,which will affect the performance of the model.In addition,annotated lymph node metastasis datasets are very scarce,which limits the performance of deep learning models.On the other hand,traditional neural networks are mostly designed for natural images and large resolution images,but are not as effective in detecting small targets such as lymph nodes,for example,classical ”U” network with simple skip connection and direct concat of the underlying features without transformation may suffer from non-linearity and thus affect the model performance.To address the above problems,this paper proposes a method for abdominal lymph node detection and metastasis prediction based on CT images.Firstly,an attention mechanismbased method for generating candidate regions of abdominal lymph nodes and a multiscale 2.5D lymph node degradation false positive method are proposed,enabling accurate detection of abdominal lymph nodes.Then a lymph node metastasis prediction model based on pseudo-labeling technology was proposed,and finally,an automatic detection and metastasis prediction system for abdominal lymph nodes was developed based on the three outcomes of the above research.(1)This paper proposes an attention-based method for generating candidate regions of abdominal lymph nodes.A ”magnifying glass” style multi-scale image is used as the input,and a residual attention skip-connection module is designed to fuse the information at each scale by transforming the multi-resolution features,thus effectively solving the problem of false detection and missed detection caused by the adhesion of lymph nodes to surrounding tissues.Meanwhile,a loss function based on pixel screening is also designed,and the constrained model focuses on the learning of lymph node features,Addressing the problem of easy loss of features for small target feature learning,The detection rate of abdominal lymph nodes is improved.The final model achieved a detection rate of 81.1% with 21 FP/case in the experiment.(2)This paper proposes a multi-scale 2.5D lymph node False Positive Reduction method.A new 2.5D input is designed to extract slices from multiple axial planes(coronal,sagittal and cross-sectional)to construct the model input.A multi-scale fusion approach is then used,allowing the model to exploit the 3D spatial information of medical images while being able to cope with the varying sizes and morphologies of lymph nodes.This results in a great increase in detection efficiency without compromising the performance of the model,enabling rapid false positive reduction for lymph node detection.Ultimately,the model achieved an AUC value of 0.94 in the experiment,reaching a 3D model performance and achieving significant improvements in time efficiency.(3)For the lymph node metastasis prediction task,this paper proposes a multi-level lymph node metastasis prediction model based on the pseudo-labelling technique,with a hierarchical idea of dealing with different volumes of lymph nodes and extended random sampling to enhance image richness.The network structure is redesigned based on Dense Net,and multiple branches are added in the feature extraction part to combine feature information of different resolutions.In order to utilise the large amount of lymph node image data without metastasis annotation,a semi-supervised learning method based on pseudo-labelling was designed to learn richer lymph node image features and enhance the feature extraction capability of the model.Ultimately this method achieved an accuracy rate of 91.52%,outperforming other advanced lymph node metastasis prediction models.(4)Based on the abdominal lymph node detection and metastasis prediction algorithm proposed in this paper,an automatic detection and metastasis prediction system for abdominal lymph nodes based on CT images is designed and implemented.The system supports multiple medical image data formats and implements core functions such as data import and parsing,candidate generation,false positive reduction,metastasis prediction and result storage.The system automates the whole process of detection and metastasis prediction from imported data to output reports to provide doctors with supplementary diagnostic advice and improves their efficiency.In summary,this paper focuses on the detection and metastasis prediction of abdominal lymph nodes.In the abdominal lymph node detection task,experiments were conducted using the CTLymph dataset and achieved a lymph node detection rate of 73.5%.In the lymph node metastasis prediction task,this paper’s method achieved an accuracy rate of 91.52%,which is superior to other state-of-the-art methods,demonstrating the validity of this study.