Improving detection accuracy of heterogeneity in biological tissues through the combination of modulation-demodulation frame accumulation techniques and enhanced vgg16

Light source has obvious absorption and scattering effects during the transmission process of biological tissues, making it difficult to identify heterogeneities in multi-spectral images. This paper achieves a gradual improvement in the classification accuracy of heterogeneities on multi-spectral transmission images (MTI) through the combination of modulation-demodulation frame accumulation (M_D-FA) techniques and enhanced Visual Geometry Group 16 (VGG16) models. Firstly, experiments are designed to collect MTI of phantoms. Then, the image is preprocessed by different combinations of frame accumulation (FA) and modulation and demodulation (M_D) techniques. Finally, multi-spectral fusion pseudo-color images obtained from U-Net semantic segmentation are inputted into the original and enhanced VGG16 network models for heterogeneous classification. The experimental results show that: While both FA and M_D significantly improved the image quality individually, their combination (M_D-FA) proved superior, yielding the highest signal-to-noise ratio (SNR) and the most accurate heterogeneous classification. Compared to the original VGG16 model, the enhanced VGG16 models gradually improved the classification accuracy. Most importantly, the 3.5 Hz M_D-FA images processed by the Visual Geometry Group 16-Batch Normalization-Squeeze and Excitation-Global Average Pooling (VGG16_BN_SE_GAP) model achieved the highest classification accuracy of 97.57%, significantly outperforming results using FA or M_D alone. In summary, this paper utilizes different combinations of FA and M_D techniques to further improve the accuracy of deep learning networks on multi-spectral images heterogeneous classification, which promotes the clinical application of multi-spectral transmission imaging technology in early breast cancer detection.