Photoacoustic/ultrasonic dual-modal endoscopic imaging enables high-penetration depth molecular imaging in biological cavities, such as the digestive tract, showing promising clinical applications. However, the compact structure of the miniature endoscopic probe makes it susceptible to electromagnetic interference (EMI), significantly affecting its imaging sensitivity and resolution. Currently, the most effective solution is to reduce EMI through averaging, but this approach incurs significant time overhead and sacrifices the system’s temporal resolution, limiting the functional imaging capabilities of photoacoustic imaging technology. Therefore, we propose two neural network-based denoising models, Aline-CNN and Aline-UNet, designed for photoacoustic waveform data, which can effectively remove electromagnetic interference from the acoustic waveform signals collected by photoacoustic/ultrasonic endoscopy (PAE/USE). Test results based on simulated photoacoustic images show that the structural similarity index (SSIM) between the denoised photoacoustic images processed by Aline-CNN and Aline-UNet and the original simulated images (ground truth) is 0.9515 and 0.9569, respectively. Furthermore, these models perform excellently on in vivo rabbit rectum imaging data. Compared to reference images, the SSIM values are 0.8239 and 0.8589, respectively. This novel denoising method can significantly reduce the number of acquisitions required and shorten the data acquisition time, thus advancing the clinical development and application of photoacoustic/ultrasonic endoscopic technology.

We propose deep learning-based denoising models, Aline-CNN and Aline-UNet, which utilize PAE/USE acoustic waveform data as input. Compared to denoising networks that use B-mode images as input, the proposed model strategy extracts more effective features from a smaller dataset, thus improving the performance of the final trained model. In addition, Aline-CNN, as a lightweight version of Aline-UNet, maintains high denoising performance while offering the potential for quick integration into the front-end of an acquisition system, improving data acquisition speeds. In this paper, we compare the proposed models with traditional denoising methods, such as wavelet filtering and singular value decomposition, as well as conventional two-dimensional networks like UNet-2D and Wave-UNet. The models’ performances are evaluated using metrics such as structural similarity (SSIM) and peak signal-to-noise ratio (PSNR). Moreover, we also train and test the models on in vivo PAE/USE data from rabbit rectums.

Using pure simulated lead-core photoacoustic data as the ground truth, we perform denoising using wavelet transform, singular value decomposition, Unet-2D, Wave-UNet, Aline-UNet, and Aline-CNN after adding the system’s pure noise signal. Among these methods, wavelet filtering effectively suppresses background noise but requires manual adjustment of filtering thresholds. In addition, it struggles to effectively suppress signals in the noise that are similar in amplitude and frequency to the effective signal, which may result in loss of useful information during the filtering process. Singular value decomposition significantly suppresses background noise while preserving signal details. However, selecting the singular value threshold still requires manual intervention, and the data volume for denoising is large, with high memory overhead. Furthermore, some noise that overlaps with the effective signal is difficult to suppress. In contrast to traditional denoising algorithms, the four deep learning models tested in this paper can effectively suppress or even eliminate background noise that does not overlap with the signal of interest. Nearly all background noise that does not overlap with the signal of interest is removed, demonstrating the considerable potential of deep learning in medical image denoising. Among the four models, UNet-2D, as a two-dimensional denoising network, appears more natural in two-dimensional visualization and preserves better structural information than the one-dimensional models, but it loses more amplitude information. Wave-UNet, Aline-UNet, and Aline-CNN perform similarly, with better preservation of amplitude information compared to the reference image. However, compared to Wave-UNet and Aline-UNet, Aline-CNN shows slightly rougher edge details of the lead core, and all three models lose details in areas where high-amplitude noise overlaps with the signal. Notably, under the same configuration conditions, the one-dimensional models (Aline-UNet and Aline-CNN) achieve better suppression of noise signals than the two-dimensional model, UNet-2D, with significantly lower memory and time costs. Specifically, Aline-UNet’s memory overhead is only 0.5% of Wave-UNet’s, while its training time is reduced from 136 s to 49 s, maintaining denoising performance. Aline-CNN, with a more lightweight structure, has lower memory and time overheads, sacrificing only slight reductions in denoising performance, making it suitable for real-time deployment in frontend systems. In in vivo PAE/USE experiments on rabbit rectums, using an average of 60 images as the ground truth, Aline-CNN and Aline-UNet demonstrate significant advantages in preserving image details in PAE, clearly restoring the distribution of rectal wall blood vessels and displaying the continuous structure of bones and tendons surrounding rectal tissues in USE.

In this paper, we introduce improved Aline-CNN and Aline-UNet model-based denoising algorithms designed to remove electromagnetic interference during PAE/USE system acquisition. While CNN and UNet networks have been used for denoising photoacoustic B-mode images, this work is the first to suppress electromagnetic noise in the Aline time-domain waveform data from the PAE/USE system. Both simulation and in vivo experimental data demonstrate the denoising capabilities of the two models. In addition, we compare the electromagnetic noise interference removal capabilities of Aline-CNN and Aline-UNet. The results indicate that, after training the network with photoacoustic and ultrasound data collected by the experimental system and using SSIM as the evaluation metric, Aline-UNet slightly outperforms Aline-CNN in denoising, but Aline-CNN offers a more lightweight structure and faster processing speed. Furthermore, the proposed models effectively suppress electromagnetic interference in photoacoustic and ultrasound data with lower training costs. They eliminate the need for multiple averaging to suppress EMI, thus reducing PAE/USE acquisition time and enhancing the clinical value and development prospects of PAE/USE endoscopy in life science research and clinical diagnosis. These models also provide insights for removing electromagnetic interference in other photoacoustic systems. However, although deep learning methods have demonstrated effectiveness in removing electromagnetic interference in the PAE/USE system, some limitations remain. The proposed models have only been validated in a limited number of experiments, and their generalization ability across different imaging algorithms and systems has not been fully explored. Future work will involve acquiring more datasets to further enhance the models’ generalization capabilities.