Computer Aided Detection Algorithm For Pulmonary Nodules Based On Deep Learning

Lung cancer is the leading cause of cancer deaths worldwide,accounting for 20% of all cancer deaths and posing a serious threat to human life and health.Clinical practice has demonstrated that early detection of lung cancer can greatly improve the survival rate of patients.Pulmonary nodules’ size and shape serve as essential criteria for early lung cancer screening,and pulmonary nodule screening plays a crucial role in lung cancer diagnosis.Low-dose spiral CT,with its advantages of fast acquisition speed,low cost,and low radiation,has been widely used in diagnosing pulmonary nodules.However,traditional manual reading methods consume radiologists’ time and energy,greatly reducing diagnostic efficiency.With the development of computer-aided detection technology,computer-aided detection systems have been applied to assist radiologists in diagnosis.The system consists of two stages: candidate nodule detection and lung nodule recognition and classification.The goal is to provide doctors with a preliminary diagnosis result through computer-aided diagnosis,which can be used as a reference for their final diagnosis.The widespread use of computer-aided detection systems has greatly reduced radiologists’ misdiagnosis and missed diagnosis rates,improving the efficiency and accuracy of lung nodule detection.Therefore,the development of computer-aided detection systems has significant significance for early screening of pulmonary nodules.This article adopts relevant deep learning methods to conduct in-depth research on lung nodule computer-aided detection algorithms and improve the shortcomings of current methods.It enhances the comprehensive detection indicators of the model.The main research content of this article is as follows:(1)This article proposes a candidate nodule detection algorithm based on 3DRetinaNet,which utilizes a sliding window method to detect candidate nodules in CT images.This article adds the Strengthen Coordinate Attention(SCA)module to the backbone network,which enhances the feature extraction ability.By fully utilizing the spherical morphological features of pulmonary nodules,this article proposes an Anchor-based sphere Io U loss function called SIouxLoss.The algorithm performs ten-fold cross-verification on the LUNA16 dataset,and the average recall CPM of candidate nodule detection task reaches 0.94.(2)This article proposes a false positive reduction algorithm for lung nodules based on cross-attention fusion mechanism.The algorithm uses a multi-scale feature refinement and fusion module and integrates features maps from different levels using cross-attention fusion mechanism to improve classification accuracy.This article improves the classifier and loss function in the training process by using a long-tail learning two-stage decoupling method,enhancing the model’s ability to classify imbalanced samples.This article also obtaines a set of effective linear weighted weights that can effectively fuse confidence sequences at different scales,further improving the classification performance of the model.The algorithm performs ten-fold cross-verification on the LUNA16 dataset,and the average recall CPM of the false-positive reduction task reaches 0.848.(3)This article proposes a lung nodule qualitative classification algorithm based on multi-dimensional fusion,which introduces High-Level Soft Activation Mapping(HESAM)module to enhance the model’s classification ability.In addition,a balanced covariance loss suitable for lung nodule qualitative classification training is introduced.This article uses a 2D convolutional neural network for single-slice classification and fused the 3D lung nodule classification results for decision-making,integrating cross-dimensional information from 2D and 3D.The algorithm can be combined with lung nodule detection to jointly determine the shape category and diameter size for benign and malignant diagnosis.The algorithm is verified on private clinical data Private Lung,and the accuracy of nodule classification reaches 0.898.