Research On Nasopharyngeal Carcinoma Image Lesion Recognition Based On Small Labeled Samples

The early detection and treatment of nasopharyngeal carcinoma has an important impact on improving the survival rate of patients.Accurate and reliable detection of nasopharyngeal carcinoma lesions in magnetic resonance imaging(MRI)can help doctors improve the efficiency and accuracy of diagnosis and treatment of this disease.Computer-aided detection methods based on deep learning can automatically detect the lesion area of nasopharyngeal carcinoma on the MRI image of the patient to help assess tumor progression.However,largescale annotation of MRI images is time-consuming and burdensome to the medical system,and is not feasible in the actual medical environment.Active learning methods can alleviate the dependence of machine learning algorithms on large-scale labeled data,and have also made progress in the field of natural images,but there are few studies in the field of medical image analysis.In addition,in the MRI images of nasopharyngeal carcinoma,the complex background occupies most of the space,while the volume of nasopharyngeal carcinoma itself is very small and infiltrative makes it difficult to distinguish from surrounding tissues.The recent deep learning methods have failed to achieve satisfactory nasopharyngeal carcinoma detection in MRI images.In view of the above problems,this paper carries out the following research:(1)Aiming at the problem that medical images are difficult to label on a large scale,this paper combines active learning and deep convolutional neural network model to propose a weakly supervised nasopharyngeal carcinoma detection method(Multi-window Settings Resampling-Active Object Detection,MWSR-AOD)suitable for MRI images,which can obtain better detection performance under a small amount of labeled data.Firstly,the pseudocolor version of the MRI image is generated based on the multi-window sampling method,which retains more abundant information and improves the information utilization of the image.Then,active learning and deep learning are integrated to construct an active detection model for nasopharyngeal carcinoma.The most representative image set is selected from the largescale unlabeled set for further annotation by experts using the uncertainty of the example-level image,which significantly reduces the demand for image annotation in deep networks.(2)Aiming at the problem that the existing algorithms fail to achieve satisfactory lesion detection on nasopharyngeal carcinoma MRI images,this paper proposes a detection method based on multi-window resampling and coordinate attention mechanism(Multi-window Settings Resampling-improved YOLOv7 embedded with a Coordinate Attention Mechanism,MWSR-YLCA)for the accurate detection of nasopharyngeal carcinoma lesions in MRI images.MWSR-YLCA first generates a pseudo-color version of the MRI image based on the multiwindow sampling method,retaining more abundant information.Then,the CA attention mechanism is introduced,and the attention convolution module MP-CA is constructed for attention feature fusion.Based on the improved YOLOv7 framework with embedded attention mechanism,a new network YLCA is constructed to detect the lesion area of nasopharyngeal carcinoma with high accuracy.To evaluate the effectiveness of the above algorithms,each algorithm was validated on an MRI image set for nasopharyngeal carcinoma.The experimental results of MWSR-AOD algorithm show that the resampling method based on multi-window setting can improve the performance of the classical deep detection model by 1.5%,while the active detection model of nasopharyngeal carcinoma only uses 20 % labeled data,which can reach 92.6 % of the performance of the deep learning detector trained with all samples.Good performance is obtained when the label set is small.Our nasopharyngeal carcinoma active detection method can detect nasopharyngeal carcinoma lesions with high accuracy without large-scale labeled data,which significantly reduces the sample labeling burden of doctors.The proposed method MWSR-YLCA was validated on the MRI image set of 800 patients with nasopharyngeal carcinoma and achieved 80.1 % m AP with only 4694 data samples.By comparing with the state-of-the-art object detection algorithms,the experimental results show that the proposed MWSR-YLCA can detect nasopharyngeal carcinoma lesions with high accuracy,and the algorithm has superior performance.