Recognition of prostate cancer based on speckle imaging combined with deep learning

Cancer is the second leading cause of death worldwide, and its diagnosis and treatment impose a significant economic burden on both society and patients. In early diagnosis, traditional pathological methods face several challenges, including high consumption of medical resources, long diagnostic cycles, and the potential for overdiagnosis, which could harm patients. To address these issues, we have developed a new method for recognizing cancerous tissues by combining speckle imaging with deep learning. Using an optical speckle imaging system, we direct a laser beam through phase screens that simulate cancerous and normal prostate tissues, obtaining speckle patterns at the receiver end that contain tissue characteristic information. These speckle patterns are then used to establish a dataset. The dataset is input into five convolutional neural networks, including CNN, AlexNet, VGG, ResNet, and Xception, for the recognition of cancerous tissues. The results show that all five convolutional neural networks can accurately recognize tissue features, achieving an accuracy of 1. To further validate the ability of convolutional neural networks to recognize tissue features in speckle patterns, the speckle patterns were cropped to 1/16 of their original size. Despite the reduced image size, four performance metrics of the CNN model—specificity, precision, recall, and F1 score—remained above 0.99. Additionally, the CNN model demonstrated advantages in terms of shorter training time and fewer training parameters compared to the other four networks. The experiments demonstrate that deep learning methods can effectively recognize cancerous tissues in optical speckle patterns.