Liu Ying, Liu Xiaoyun, Gao Siyu, Chai Tengfei, Zhao Zihao, Wang Hongwei, Jin Yumeihui, Jiang Yueqiu
School of Science, Shenyang Ligong University, Shenyang, 110159, China.
Science and Technology Department, Shenyang Ligong University, Shenyang, 110159, China.
Sci Rep. 2025 Apr 17;15(1):13292. doi: 10.1038/s41598-025-88441-2.
The identification of cancerous tissues remains challenging due to the complexity of experimental methods and low identification accuracy rates. Therefore, this paper proposes a rapid identification method. We introduce a new theoretical transmission method for modeling laser beams transport in cancerous and normal biological tissues. Using this method, laser speckle patterns carrying tissue information are obtained at the light transmission receiving plane. Then, we propose a three-task residual neural network, T-ResNet-18, for identifying speckle images. We simulate the human normal and cancerous prostate tissues, rat normal and cancerous tissues, and normal and abnormal cell suspension as training samples. The results show the identification accuracy exceeding 99%. Additionally, we discuss the impact of varying dataset sizes, training epochs, and tissue thickness on identification accuracy and compare the performance of T-ResNet-18 with ResNet-18, VGG16 and AlexNet, showing that T-ResNet-18 significantly outperforms classic neural networks.
由于实验方法的复杂性和较低的识别准确率，癌组织的识别仍然具有挑战性。因此，本文提出了一种快速识别方法。我们引入了一种新的理论传输方法，用于对激光束在癌组织和正常生物组织中的传输进行建模。使用这种方法，在光传输接收平面上获得携带组织信息的激光散斑图案。然后，我们提出了一种用于识别散斑图像的三任务残差神经网络T-ResNet-18。我们模拟了人类正常和癌性前列腺组织、大鼠正常和癌性组织以及正常和异常细胞悬液作为训练样本。结果表明识别准确率超过99%。此外，我们讨论了不同数据集大小、训练轮次和组织厚度对识别准确率的影响，并将T-ResNet-18与ResNet-18、VGG16和AlexNet的性能进行了比较，结果表明T-ResNet-18明显优于经典神经网络。
