Research On Classification And Segmentation For Abdominal Tumor Images Based On Deep Transfer Learning

The abdominal organs play an important role in the digestive,reproductive,and urinary functions of the human body.However,abdominal organs are highly variable and intricately interconnected,making the diagnosis of abdominal tumors a huge challenge.With the rapid development and wide application of artificial intelligence in healthcare,medical image analysis has become an important method for diagnosing and treating abdominal tumors.To address the problems of irregular size and shape of abdominal tumors,limited amount of tumor image data,low image contrast,and fuzzy boundaries between different soft tissues or between soft tissues and lesions,deep learning methods for medical imaging can assist doctors in accurately targeting the location of abdominal tumor images,evaluating the size of the tumors,and also help doctors to formulate a preoperative plan,guide the operation in hand,or assist in the management of the prognosis.Starting from the actual needs and problems of clinical medicine,this dissertation selects the four kinds of abdominal tumor categories that rank among the top10 in terms of incidence rate in China,including colorectal,pancreatic,liver and cervical tumors,and carries out a comprehensive analysis and in-depth study from the perspective of their tumor pathological characteristics and imaging features,and innovatively proposes a deep transfer learning-based classification and segmentation algorithm of abdominal tumors images to improve the performance of diagnosing tumors of abdominal organs.The work of this dissertation mainly includes the following parts.1.To address the problems of colorectal images with bright light specular reflection,external interference caused by abnormal exposure,and the difficulty of classifying three types of samples,this dissertation proposes a colorectal image classification model based on deep meta-differential self-paced transfer learning.The model adopts the idea of parameter transfer learning for different scenarios.It contains two modules,namely,the module of colorectal lesion highlight detection and removal and the module of meta-differential self-paced learning classification.This model firstly uses computer vision and deep transfer learning methods for colorectal image highlight detection,followed by transfer learning methods based on bright specular natural image dataset to repair colorectal image data.On this basis,this dissertation designs a meta-differential self-paced learning strategy suitable for colon image classification.The strategy is used during the training process to continuously improve the ability of collaboration among the age parameter,model parameter,and sample weight parameter,and adaptively learn the importance of each sample to improve the performance of colorectal classification model.Compared to the original dataset,the classification accuracy of the clinical colorectal dataset based on meta-differential self-paced learning strategy with highlight reparation is improved by about 9% to 11% on different classification network models.The experimental results confirm the advantages of the deep meta-differential self-paced transfer learning method for the problem of optimizing colorectal classification networks with improved sample classification performance.2.To deal with the high precision requirements for pancreatic tumors during radiotherapy,as well as the problems that uni-modality images are difficult to meet the needs of radiation therapy and the small amount of data in multi-modality images,this dissertation leverages the potential correlation information between multi-modal abdominal images and proposes a multi-modal pancreatic tumor segmentation framework based on multi-level generative transfer learning.The model utilizes the idea of associated transfer learning of different modalities and contains three sub-models,namely,a multi-modal image generation model based on multi-channel Cycle GAN with transfer learning,a PET pancreatic tumor localization model based on transfer learning,and a pancreatic tumor segmentation model based on an improved Trans Unet segmentation network.These three sub-models associate the lung tumor PET/CT image public dataset,the pancreatic tumor CT image public dataset,and the clinical pancreatic tumor PET/MR image dataset closely,and finally obtain clinical pancreatic tumor PET/MR image segmentation results.The proposed method is compared with other segmentation methods on clinical PET/MR image datasets of pancreatic tumors,and the results of multimodal pancreatic tumor segmentation are significantly improved.Compared with other methods,this method improves the Dice Similarity Coefficient of pancreatic tumor segmentation on clinical pancreatic tumor PET/MR image datasets by about 1.7%.The experimental results confirm the effectiveness of the method in mining potential correlation information between different multimodal image datasets and dealing with the pancreatic tumor segmentation problem.3.To tackle the problems of the complexity and variability of liver tumors,the blurring of the boundaries between tumor regions and normal tissues,and the fact that it is usually more difficult to acquire multi-modal imaging data in a clinical environment,this dissertation proposes a liver tumor segmentation model based on multi-phase awareness and knowledge distillation in the missing-phase situation.The model applies the idea of model transfer learning in different phases,which contains three components,namely,multi-phase feature extraction and fusion perception mechanism,a knowledge distillation liver tumor segmentation model based on Eff-Unet network,and transfer strategy for missing multi-phase image data.Using the perceptual information of arterial-phase liver tumor enhanced CT images,the knowledge distillation method is used to assist in improving the segmentation task of venous-phase liver tumor enhanced CT images.In addition,the models acquired by the multi-phase enhanced CT image segmentation network are transferred to the single-phase enhanced CT image segmentation network in a phase-missing scenario to achieve the single-phase liver tumor enhanced CT image segmentation task.Compared with other methods,this method improves the Dice Similarity Coefficient of liver tumor segmentation by 3% on a single-phase liver tumour dataset under phase-missing scenarios.The experimental results confirm the advantages of the liver tumour segmentation method with multi-phase awareness and knowledge distillation.4.In view of the problems of small cervical tumor volume,fuzzy boundary,inhomogeneous internal grayscale,and that the thicker layers of magnetic resonance scanning result in axial,coronal and sagittal images presenting different information,a cervical tumor segmentation method based on multi-view feature transfer learning is proposed.The method adopts the idea of viewpoint transfer learning in different dimensions,which mainly includes three aspects,namely,multi-scale residuals and multi-scale bottleneck attention mechanism,multi-viewpoint feature transfer learning strategy,and the construction of cervical tumor segmentation loss function.The segmentation model employs a dual branch network model with 2D and 3D perspectives.This method constructs a 2D cervical tumor axial slice encoder-decoder model,which is used to achieve multi-view feature transfer learning from 2D features to a 3D segmentation network to extract the global position and shape information of the cervical tumors efficiently.Multi-scale residual module with multi-scale residual attention module is employed in 3D cervical tumor segmentation network to capture features at different scales and recalibrate the importance of features at each scale to avoid information redundancy.This method is compared with other segmentation methods on a clinical MR image dataset of cervical tumors,and the cervical tumor segmentation results are significantly improved.Compared with other methods,this method improves the Dice Similarity Coefficient of cervical tumor segmentation by 4.1% on the clinical cervical tumor MR image dataset.The experimental results confirm the effectiveness of viewpoint transfer learning strategy,multi-scale residuals and attention mechanism in dealing with the cervical tumor segmentation problem.