3DNET: Deep Vein Thrombosis Detection via Whale Social Spider Optimized Dense Dilated CNN with Dual Images

Deep vein thrombosis (DVT) is a widespread venous disorder characterized by the formation of blood clots within the deep veins, most commonly in the lower extremities. The existing methods rely only on Duplex ultrasound images, which often struggle with noise interference, limited contrast, and insufficient visibility of thrombus boundaries. Moreover, single modality ultrasound analysis lacks sufficient detail, which leads to misclassification of thrombus conditions when there are unclear boundaries or variations in shape and location. To overcome these challenges, a novel 3DNet has been proposed that utilizes both Duplex and Doppler ultrasound images for more accurate DVT classification. A two-stage coarse-to-fine thrombus segmentation neural network is employed to precisely delineate thrombus boundaries and vein structures by ensuring accurate localization of pathological regions. Efficient channel attention-embedded RegNet (ECA-RegNet) is used to extract structural features by strengthening the discriminative representations. Whale social spider optimization (WSSO) is employed to select the most relevant features by effectively eliminating redundant and irrelevant attributes for accurate thrombus classification. The dense-dilated convolutional neural network (CNN) is used to categorize into coronary thrombosis (CT), venous thromboembolism (VTE), and pulmonary embolism (PE), enabling reliable and clinically applicable DVT stage identification. The proposed 3DNet attains an overall accuracy of 99.13% based on the dual images. The proposed 3DNet enhances the total accuracy by 9.7%, 9.65%, 18.8%, and 15.8% higher than CNN-Mobilenet, KNN, SVM, and Fine-tuned ResNet, respectively. The proposed 2-stage coarse-to-fine thrombus segmentation neural network increases the overall DI by 22.13%, 11.34%, and 6.31% better than U-Net, V-Net, and SegNet, respectively.