.png){kind=icon}
.png){kind=icon}
Return
Cancer Tissue Recognition Based on Convolutional Neural Network
.png){kind=icon}
PRE
AI
H
X
T
Y
J
T
Y
DOI：10.3788/LOP252131.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective This study aims to develop an efficient artificial intelligence method for early cancer diagnosis by combining vortex beam speckle imaging with deep learning. Traditional pathological diagnosis relies on manual microscopic examination of histological images, which is time-consuming, inefficient, and may delay treatment. Although traditional optical imaging is non-invasive, its utility is limited by strong scattering and absorption effects in biological tissues, leading to wavefront distortion and resolution reduction. Vortex beams, carrying orbital angular momentum (OAM) and exhibiting a helical phase structure [spiral phase factor is exp (il theta)], generate speckle patterns rich in tissue-specific optical information, offering a novel approach for pathological analysis. We propose a convolutional neural network model based on a residual structure for analyzing such speckle images. This model can simultaneously achieve precise identification of cancerous tissues and extract key optical parameters closely related to scattering characteristics-fractal dimension and autocorrelation length-thereby revealing changes in the tissue microstructure. This multitask learning framework enhances the robustness and interpretability of the diagnosis. The proposed method provides a rapid, non-invasive, and high-precision cancer screening tool, demonstrating the effective integration of computational optics and deep learning in medical diagnostics and showing promising clinical application prospects. Methods We propose a residual-structured neural network, termed ResHDCNet, to simultaneously identify cancerous tissues, extract scattering-related parameters-fractal dimension (Df) and autocorrelation length (lc)-from speckle images. The ResHDCNet architecture comprises 26 convolutional layers and 6 fully connected layers, with modified residual blocks adopting a batch normalization (BN)-activation-convolution order to stabilize feature learning. A hollow dense connection (HDC) module employing dilated convolutions (r=1, 2, 5) is incorporated to expand the receptive field and capture multi-scale information with minimal computational overhead. The model processes single-channel speckle images through a downsampling convolution-pooling stage, followed by hierarchical convolution, global average pooling, and three parallel output heads for classification and regression tasks. A dataset with 8400 speckle images is generated using phase screen simulations of six types of normal and cancerous tissues. Results and Discussions ResHDCNet achieves accuracies of 97.74 degrees 0, 97.50 degrees 0, and 97.74 degrees 0 for cancer tissue classification, Df prediction, and lc prediction, respectively. Specificity exceeds 99.50 degrees 0, while recall and F1-score are no lower than 97.70 degrees 0. Under dataset pruning experiments, where images are cropped to 90 degrees 0, 80 degrees 0, 70 degrees 0, 60 degrees 0, and 50 degrees 0 of their original size and then upsampled, the model maintains accuracies above 93.40 degrees 0, demonstrating robust feature extraction under partial information loss. Comparative analysis further shows that ResHDCNet outperforms ResNet18, ResNet34, VGG16, AlexNet, and MobileNetV4 across all evaluation metrics, with accuracy, specificity, recall, precision, and F1-score consistently above 97.70 degrees 0. Conclusions This work introduces ResHDCNet, a residual-structured deep learning framework that integrates vortex beam speckle imaging with optical scattering models for cancer diagnosis. The experimental results demonstrate its superior capability in both tissue classification and optical parameter estimation, even under constrained data conditions. By leveraging the unique advantages of vortex beams in deep-tissue imaging, this method provides a rapid, non-invasive, and high-precision tool for cancer screening, highlighting the promising synergy between computational optics and deep learning in clinical diagnostics.
Keywords:
vortex beam
convolutional neural network
biological tissue phase screen
speckle recognition
Journal
L
IF:
1
Papers:
505
Citations:
0