Research On The Segmentation Of New Coronary Pneumonia Lesions Based On Transfer Learning

Computed tomography(CT)imaging technology plays a crucial role in diagnostic tasks such as medical image detection,segmentation,and classification,providing strong evidence for physicians in clinical diagnosis,especially during the epidemic of Corona Virus Disease(COVID-19)in the past two years,the excellent performance of artificial intelligence technology(AI)in CT imaging has contributed to COVID-19's rapid containment.That is,the use of AI deep learning technology to automatically and precisely segment infected lesion areas from CT images will help physicians to diagnose and treat more accurately.However,deep learning techniques often have the following challenges in the task of COVID-19 lesion segmentation: 1)weak and irregular tissue boundary information.the COVID-19 lesions on CT images appear more like ground glass,resulting in blurred lesion boundaries with the surrounding tissues and organs;2)small data problems in medical images due to the early stage of the disease.The sudden outbreak of the disease leads to the lack of labeled data that can be used for deep learning tasks,and the small data problem limits the training of the model;3)the problem of no labeling of data due to condition restrictions.accurate labeling data of COVID-19 lesions often requires experts to complete,which is time-consuming and labor-intensive,resulting in a large amount of data without accurate pixel-level labeling.To address the above challenges,this paper proposes a series of deep learning methods and techniques to solve these problems.There are three main aspects: 1)a convolution neural network model with an improved residual module as the component backbone network,while multiple global context-sensitive field feature extraction modules as the bridge,and a feature fusion module that fuses channel attention and spatial attention mechanisms in a parallel manner as the decoder is designed to accomplish the task of lesion segmentation on CT images as a solution to the first challenge problem;2)a two-stage transfer learning technique for cross-domain data to solve the small data problem of medical images caused by the early stage of COVID-19;3)for a large amount of existing unlabeled data,a semi-supervised learning strategy is designed to dynamically generate pseudo-labels for unlabeled data,and a transfer learning strategy of pre-training pseudo-labeled data with fine-tuning real labeled data is used to improve the performance of the model.To validate the segmentation performance of the proposed model in this paper,quantitative and qualitative comparisons are made with five classical medical image segmentation models on the Mos Med Data dataset,and the method in this paper achieves excellent results on six medical image segmentation evaluation metrics,namely,DSC,SEN,SPE,PPV,VOE and RVD,respectively.Meanwhile,by designing transfer learning strategies with different steps,the results on DSC show that two-stage cross-domain transfer learning improves 1.63% over one-stage transfer,and 4.32%over no transfer at all,while all other metrics are improved by a certain magnitude,proving that two-stage cross-domain transfer can effectively alleviate the problem of small data in medical images.Finally,by designing a series of training strategies for the pseudo-labeled data generated under semi-supervised learning,it is shown that the model segmentation performance is improved by 5.93%,5.77%,7.91%,and 8.05% on DSC,SEN,PPV,and VOE,respectively,under the model pre-training using pseudo-labeled data and the fine-tuning strategy for real labeled data,demonstrating that the semi-supervised learning algorithm can not only improve the model segmentation performance,but also can effectively solve the problem of unlabeled data.