Enterprise AI Analysis
Computational self-interference spectroscopy with reconstruction enhanced by CNN for quantitative label-free imaging
Authors: Jing Gao*, Hao Xu, Zhidong Han
Publication: CCEAI 2026, January 30-February 01, 2026, Hong Kong, Hong Kong
Label-free quantitative phase imaging (QPI) enables nanoscale visualization of sample structure and dynamics. However, conventional QPI methods often depend on complex interferometric setups. While coherent white-light spectroscopy simplifies thickness and refractive index measurements, its efficacy is compromised by limited signal-to-noise ratio (SNR) and chromatic aberrations. To address these limitations, we present a computational self-interference spectroscopy framework augmented with a convolutional neural network (CNN). Validation on silica phantoms demonstrates a 12.5-fold SNR enhancement, a height sensitivity improvement from 12.4 nm to 1.1 nm, and a lateral resolution of 592.4 nm, effectively suppressing imaging artifacts. The practical utility of our method is further confirmed through high-fidelity quantitative imaging of living cells.
Executive Impact: Key Findings
This analysis reveals critical performance enhancements achieved by integrating Convolutional Neural Networks (CNNs) into computational self-interference spectroscopy. The breakthrough offers superior imaging fidelity and resolution for label-free quantitative phase imaging.
0
SNR Enhancement
0
Height Sensitivity
0
Lateral Resolution
Key Takeaways for Enterprise AI Adoption:
- ✓ **Enhanced Imaging Accuracy:** CNN integration significantly boosts Signal-to-Noise Ratio and minimizes chromatic aberrations, leading to more precise quantitative phase imaging, critical for high-stakes analytical applications.
- ✓ **Superior Resolution & Detail:** The framework improves lateral resolution and height sensitivity, enabling visualization of nanoscale structures and dynamic morphological changes with unprecedented clarity.
- ✓ **Label-Free & Non-Invasive:** Offers a powerful alternative to traditional methods by eliminating the need for exogenous fluorescent labels, reducing phototoxicity and instability, ideal for live-cell studies and sensitive biological samples.
- ✓ **Scalable & Robust Framework:** The hybrid approach, combining physics-motivated spectroscopy with data-driven CNN reconstruction, provides a practical and scalable solution for high-fidelity imaging across diverse fields.
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Introduction
Label-free imaging with nanoscale sensitivity is critical for understanding the structure and dynamic morphology of both biological and non-biological samples [1-3]. While mainstays such as phase contrast and differential interference contrast (DIC) microscopy offer valuable qualitative insights, they inherently lack the quantitative rigor required for accurate nanoscale topological reconstruction. To address these quantitative limitations, interferometry-based methodologies have become indispensable. Furthermore, avoiding the phototoxicity and instability often associated with exogenous fluorescent labels remains a priority for live-cell studies [4]. Consequently, Quantitative Phase Imaging (QPI) has gained significant traction as a powerful alternative. By employing interferometric configurations that modulate either the illumination wavelength or the coherent phase, QPI facilitates the precise derivation of sample thickness distributions [5-7]. Owing to its distinct advantages in system integration, operational simplicity, and high measurement fidelity, QPI has been extensively adopted in fields ranging from label-free biosensing to intricate three-dimensional (3D) topographical mapping [8-11]. Nevertheless, wavelength-scanning QPI techniques face specific challenges compared to phase-shifting approaches. Primarily, they are more susceptible to imaging noise, which can severely compromise reconstruction accuracy. Additionally, given that optical transfer functions exhibit wavelength-dependent variations, even minor spectral fluctuations can impede precise structural recovery. These combined factors-limited Signal-to-Noise Ratio (SNR) and chromatic inconsistencies-currently constrain the broader utility of wavelength-resolved QPI. Enhancing the fidelity of wavelength-scanning QPI systems necessitates overcoming two simultaneous hurdles: boosting the Signal-to-Noise Ratio (SNR) to effectively recover faint signals from background interference, and securing apochromatic uniformity to maintain spatial consistency across the diverse spectral channels used for reconstruction. Alongside developments in optical hardware, Deep Learning-specifically Convolutional Neural Networks (CNNs)-has emerged as a premier computational strategy for advanced image processing. The core strength of these networks lies in their capacity to autonomously extract multi-level feature hierarchies from raw input data, a capability that has revolutionized the field of computer vision [12-14]. While early architectures like LeNet-5 established the foundational principles [15], it was the transformative success of AlexNet in large-scale recognition tasks that definitively marked the onset of the modern deep learning era [16]. Currently, CNN-driven algorithms are extensively applied in diverse domains, showing particular efficacy in tasks such as image denoising and super-resolution restoration [17-19]. Given their flexibility in processing single-input images under varying illumination conditions, these data-driven models present a viable pathway for overcoming the limitations of traditional QPI. Beyond performance enhancement, it is essential to emphasize that the proposed framework is not a generic deep-learning-assisted imaging pipeline, but a physically motivated integration of self-interference spectroscopy and data-driven reconstruction. In wavelength-scanning QPI, the intrinsic coupling between spectral dispersion, coherence-induced noise, and spatial resolution degradation renders conventional post-processing strategies insufficient. Purely optical compensation approaches are often constrained by system complexity and alignment sensitivity, whereas purely data-driven methods lack physical interpretability and robustness across spectral channels. By embedding CNN-based preprocessing explicitly prior to interferometric reconstruction, the proposed framework establishes a hybrid paradigm in which deep learning serves to regularize wavelength-dependent distortions while preserving the underlying physical interference model. This synergistic design ensures that quantitative reconstruction accuracy is improved without violating optical constraints, thereby offering a practical and scalable solution for high-fidelity label-free imaging.
Methods and Materials
To rigorously validate the proposed framework, two distinct specimen categories were utilized. Initially, for biological assessment, HeLa cells (ATCC, catalog no. EY-X0129) were cultivated on standard glass substrates sourced from Xinlan Company, as illustrated in Fig. 1A. Subsequently, we employed custom-engineered silica square phantoms with precise dimensions (5 µm side length and 500 nm thickness) to serve as a ground truth. While the raw silica material was procured from Shunshen Company (Zhuhai, China), the micro-fabrication of the square patterns was executed by the Institute of Microelectronics, Chinese Academy of Sciences (Beijing, China) [20]. To verify the surface cleanliness and structural integrity of these silica samples, we performed high-resolution inspection using a Scanning Electron Microscope (SEM, Model: FEI Helios NanoLab 600i, USA), as depicted in Figure 1B. Data acquisition was performed using an epi-confocal microscope system manufactured by Nikon (Model: C1/C1si, Japan). The optical setup incorporated three distinct laser excitation sources (operating at 405 nm, 532 nm, and 647 nm). To facilitate quantitative reconstruction, these specific wavelengths were individually employed to capture separate confocal image stacks. Furthermore, specific dichroic beam splitters (405/532/647 nm, sourced from Yokogawa CSU, Japan) were installed to filter ambient light, effectively constraining the illumination bandwidth to within 10 nm for each respective channel. To capture confocal bright-field coherent imagery, the imaging plane was kept free of additional filters. Following established calibration protocols, we implemented a speckle noise suppression strategy by recording the background illumination field using standard glass substrates. The raw coherent images were subsequently normalized by these illumination patterns to yield calibrated reflectance data. For high-fidelity signal acquisition, particularly critical for quantitative analysis, the system was coupled with a high-sensitivity Andor iXon EMCCD detector (Model: DU897U). Moving to the data curation phase, the preparation workflow involved three distinct stages. Initially, to train the denoising CNN model, we compiled a comprehensive dataset comprising 1,500 scans of the silica phantoms for each spectral channel. While the original acquisitions were stored at a resolution of 512x512 pixels, these were systematically segmented (patched) into smaller 32×32 pixel blocks. This augmentation strategy resulted in a total dataset of 72,000 patches-allocating 24,000 distinct samples for each wavelength channel-to ensure robust network convergence. Since the designed pattern of the silica samples was determined, the noises could be recognized apparently. The reflectance indexes were set to be one in the image backgrounds, and the ones of the silica squares were set to be the theoretical values according to Equation 1. The initial reflectance imaging results were utilized as the inputs of the training sets, while the processed reflectance imaging results as the outputs. To ensure rigorous evaluation and reproducibility, the entire dataset was randomly partitioned into three subsets: 80% for training (57,600 patches), 10% for validation (7,200 patches), and 10% for testing (7,200 patches). The testing set was strictly isolated during the training phase to prevent data leakage and ensure unbiased performance assessment. In the second phase of data curation, we focused on training the resolution enhancement model. We utilized the same extensive dataset of 72,000 segmented patches (32×32 pixels) generated in the denoising step. The strategy for generating ground-truth data leveraged the physical property that the 405 nm illumination channel intrinsically offers superior lateral resolution due to its shorter wavelength. Consequently, we adopted the 405 nm reflectance topology as the structural reference. We computationally remapped the intensity profiles of the high-resolution 405 nm images to match the theoretical reflectance coefficients of the 532 nm and 647 nm wavelengths. These synthetically aligned, high-resolution images served as the target outputs (ground truth) for training, while the raw, experimentally acquired three-channel images served as the inputs. To empirically validate the framework's utility in biomedical settings, biological specimens were processed using a standardized fixation protocol. Once the HeLa cells achieved optimal adhesion to the glass substrate, the samples were removed from the culture environment and submerged in a 4% paraformaldehyde fixative (sourced from Bogu Company, China) for a precise duration of 15 minutes. To eliminate potential chemical artifacts, the fixed samples underwent a rigorous two-step rinsing procedure with deionized water, followed by careful desiccation using a low-pressure nitrogen gas stream to mitigate water spotting. For the final quantitative imaging demonstration, we specifically targeted a HeLa cell undergoing cytokinesis (cell division) to challenge the system's ability to resolve dynamic morphological changes.
Results and Discussion
In this experiment, we utilized the silica phantom samples to prove the quantitative denoising performance of the CNN model. The experiment results are shown in Figure 4. From 4A-4C, the original reflectance images of the three channels were illuminated with the wavelength of 405 nm, 532 nm and 647 nm, respectively. From 4D-4F, the denoised reflectance images of the three channels were the corresponding outputs of the CNN model. Peak signal-to-noise ratio (PSNR), which in commonly used to assess the denoising performance of the CNN model, was demonstrated as the following equation [22]: In this equation, the term noise represents the average noise intensity, while k denotes the image bit depth, which was configured to 16-bit for this specific study. Theoretically, a higher PSNR value correlates with reduced noise artifacts and superior image fidelity. Our quantitative analysis reveals that the raw images (Figure 4A-C) exhibited a baseline PSNR of 69.7 dB. Following the CNN processing (Figure 4D-F), the PSNR significantly increased to 80.7 dB. This substantial increment corresponds to a 12.5-fold enhancement in SNR, confirming the model's efficacy in extracting signals from high-noise backgrounds. Following the noise suppression tests, we further evaluated the achromatic capabilities of the resolution correction CNN using the same silica phantom benchmarks. The qualitative improvements are systematically presented in Figure 5. Panels 5A through 5C depict the initial reflectance maps under 405 nm, 532 nm, and 647 nm illumination, respectively, where chromatic aberration is evident. Panels 5D through 5F display the CNN-processed outputs, showing a marked improvement in structural alignment across channels. To provide a rigorous quantitative evaluation, we analyzed the intensity profiles extracted from the central axis of the phantom features. The lateral resolution for each channel was determined based on the Full Width at Half Maximum (FWHM) criterion [23]. As illustrated in the Gaussian-fitted curves of Figure 6A (raw data) and Figure 6B (processed data), the disparity in resolution is substantial. Prior to correction, the lateral resolutions were measured at 496.1 nm (405 nm), 651.7 nm (532 nm), and 792.6 nm (647 nm), indicating significant wavelength-dependent degradation. Post-processing, the resolutions converged to 496.1 nm, 513.5 nm, and 550.9 nm, respectively. Beyond the apparent FWHM improvement, the key outcome of this correction is the increased structural concordance among spectral channels. Multi-wavelength interferometric reconstruction implicitly assumes that corresponding spatial features originate from the same physical location and differ primarily in wavelength-dependent phase response. When chromatic blur and misregistration are present, this assumption is violated and the fitting procedure mixes inconsistent spatial information, leading to biased thickness/phase estimates. By improving cross-channel edge co-localization and suppressing wavelength-dependent spreading, the proposed correction step enforces the underlying model assumption required by Eq. 3), thereby reducing systematic reconstruction error rather than merely improving visual sharpness. As shown in Figure 6A, the raw Gaussian-fitted Point Spread Functions (PSFs) for the 405 nm, 532 nm, and 647 nm channels are represented by the black, blue, and red curves, respectively, while the green curve depicts the complex image fitting. The wide distribution of the raw curves in Figure 6A illustrates significant chromatic aberration. In contrast, Figure 6B presents the corresponding data after processing with the proposed CNN model. The convergence of the curves in Figure 6B visually confirms the effective correction of wavelength-dependent blur, aligning the spatial resolution across all spectral channels. Crucially, it is important to clarify the mechanism behind this enhancement. The observed improvement does not stem from the introduction of external spatial frequencies beyond the optical diffraction limit. Rather, the CNN model effectively deconvolves the chromatic blur, thereby sharpening the effective Point Spread Function (PSF). This computational "tightening" of the PSF results in a more distinct delineation of sample edges, allowing for the precise recovery of fine structural details that were previously obscured by chromatic dispersion. In Figure 7, the 3D reconstruction results of the silica sample are shown. Figure 7A shows the reconstruction results with the method proposed in this work, while Figure 7B showed without the CNN model process. The color of each pixel represented the detected thickness. With the CNN model, the PSNR value was improved from 47.2 dB to 58.2 dB, and the lateral resolution was improved from 688.9 nm to 592.4 nm. In Figure 7A, the average value of the detected thickness was 500.0 nm which equaled the actual value, and the standard deviation was 0.38 nm. In Figure 7B, the mean detected thickness was 503.8 nm, and the standard deviation was 4.15 nm. This substantial deviation in the baseline measurement is primarily attributed to the inherent coherence artifacts of the imaging system. Specifically, parasitic interferences arising from internal optical reflections and laser speckle noise introduce high-frequency fluctuations on the sample surface profile. The significant reduction in standard deviation observed in our method (Fig. 7A) confirms that the CNN module successfully disentangles these coherent artifacts from the true topological signal, rather than merely performing a low-pass smoothing operation. Here, Re is the axial resolution, o is the standard deviation and K is the mean detected thickness divided by the actual value. According to this equation, the axial resolution improved from 12.4 nm to 1.1 nm with the proposed method. To demonstrate the practical value of the proposed work, we performed imaging experiments on multiple HeLa cell samples (N=30) to verify reproducibility. A representative case of a single HeLa cell undergoing cytokinesis (dividing into two cells) was selected for detailed reconstruction analysis, as shown in Figure 8. The color of each pixel represents the detected coherent phase. Figure 8A-8C was the raw tomography reconstruction result of the z=3 µm, z=5 µm and z=7 µm respectively. Figure 8D-8F was the improved tomography reconstruction result of the corresponding layer with the proposed method. With the proposed CNN-based method, the resolution of the single cell was improved greatly, and the spindle fibers could be observed more clearly in Figure 8D-8F. This result proved that the proposed method performed better for QPI cell imaging. It is worth noting that compared to traditional model-based deconvolution methods, which rely heavily on accurate Point Spread Function (PSF) estimation and are sensitive to noise, our method demonstrates superior robustness in low-SNR environments. Furthermore, unlike generic deep learning denoising architectures (e.g., standard DnCNN or U-Net) that often treat all spectral channels uniformly, our dual-CNN framework explicitly addresses the wavelength-dependent resolution degradation. By decoupling the SNR enhancement and resolution correction tasks, the proposed method preserves high-frequency spectral details that are typically smoothed out by conventional end-to-end reconstruction networks, thereby achieving a better balance between noise suppression and structural fidelity.
12.5x
Increased Signal-to-Noise Ratio
Enterprise Process Flow
Raw Data Acquisition
→
CNN-based Denoising
→
CNN-based Resolution Correction
→
Interferometric Reconstruction
→
Quantitative Imaging Results
| Feature | Traditional QPI | CNN-Enhanced QPI |
|---|---|---|
| Setup Complexity |
|
|
| Noise Susceptibility |
|
|
| Chromatic Aberrations |
|
|
| Reconstruction Accuracy |
|
|
| Live-Cell Compatibility |
|
|
High-Fidelity Imaging of Living HeLa Cells
The framework successfully captured dynamic morphological changes during HeLa cell cytokinesis, demonstrating its robustness and practical utility for biomedical research. This high-fidelity quantitative imaging was achieved without exogenous fluorescent labels, addressing a key challenge in live-cell studies.
- Resolved dynamic morphological changes
- Enabled label-free quantitative imaging
- Demonstrated practical utility in biomedical settings
Calculate Your Potential AI Impact
Estimate the efficiency gains and cost savings your organization could achieve by implementing advanced AI solutions, tailored to your specific operational context.
Your AI Implementation Roadmap
A phased approach to integrate computational spectroscopy with CNN-enhanced reconstruction into your research or industrial workflows, ensuring seamless transition and maximized impact.
Phase 1: Needs Assessment & Data Audit
Analyze existing imaging workflows, data characteristics, and define specific quantitative imaging goals. Evaluate current hardware capabilities and identify potential data sources for CNN training.
Phase 2: System Integration & Calibration
Integrate computational spectroscopy hardware with data acquisition systems. Implement initial calibration protocols and establish a baseline for performance metrics.
Phase 3: CNN Model Customization & Training
Adapt and fine-tune CNN architectures (denoising and resolution correction) using domain-specific datasets. Optimize training parameters for robust performance and generalizability.
Phase 4: Validation & Workflow Integration
Rigorously validate the CNN-enhanced framework using phantom and biological samples. Integrate the optimized system into daily operational workflows and provide user training.
Phase 5: Performance Monitoring & Iterative Improvement
Continuously monitor imaging accuracy, resolution, and data processing efficiency. Implement feedback loops for iterative model updates and system enhancements based on real-world application.
Ready to Transform Your Imaging?
Leverage the power of AI-enhanced computational spectroscopy for unparalleled quantitative label-free imaging. Our experts are ready to design a tailored solution for your organization.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat