Fast and compact temporal convolutional network for non-resonant background removal in broadband CARS.

Broadband coherent anti-Stokes Raman scattering (BCARS) microscopy enables the acquisition of full Raman spectra (400-3200 cm-1) within milliseconds, offering substantial potential for rapid, label-free chemical imaging. However, the non-resonant background (NRB) originating from four-wave mixing causes line-shape distortion and diminishes contrast. Traditional NRB removal methods require prior knowledge of NRB spectral profile, thereby increasing experimental complexity. Although existing deep learning approaches have achieved effective NRB suppression with a maximum coefficient of determination of 0.99, they generally suffer from a trade-off between prediction accuracy and inference efficiency. Typical LSTM and GAN models contain 3841 and 55,249 trainable parameters with single-spectrum inference latencies of 42 ms and 1.6 ms respectively, severely limiting real-time BCARS imaging applications. This work proposes a lightweight temporal convolutional network (TCN) based on dilated convolution, which expands the model receptive field while maintaining low computational complexity. The proposed model achieves superior prediction performance with an R2 value greater than 0.95 and significantly reduces computational cost, realizing sub-millisecond single-spectrum prediction at a latency of 0.58 ms, which provides a novel solution for high-speed real-time BCARS imaging. Comparative experiments demonstrate that the TCN model outperforms state-of-the-art methods including SpecNet, VECTOR, LSTM, Bi-LSTM, GAN, and CNN + GRU in identifying authentic Raman peaks. Ablation studies verify that the dilated convolution architecture effectively balances structural simplicity, prediction accuracy, and computational efficiency, making it highly suitable for high-speed BCARS spectral detection and imaging tasks.