HS-QFNet: Deep learning-enhanced hyperspectral fluorescence correction for accurate in vivo photosensitizer concentration quantification

Summary

Accurate quantification of photosensitizer concentration is essential for effective fluorescence-guided surgery and personalized photodynamic therapy, but it is hindered by tissue-induced fluorescence distortions because existing correction methods have limited accuracy and clinical adaptability. We present a deep-learning-based fluorescence correction algorithm (Hyperspectral Quantitative Fluorescence Network [HS-QFNet]) that integrates hyperspectral fluorescence and diffuse reflectance image features from a phantom array with broad optical properties, combined with an attention mechanism to model nonlinear relationships between signal distortion and tissue optical properties, enabling precise detection of photosensitizer spatial distribution. Validated in phantoms, it achieved a mean absolute error (MAE) of 0.21 μM—a 68% improvement over traditional methods (0.65 μM MAE). In mouse tumor models, it maintained an MAE of 0.31 μM with a 0.957 correlation to true concentration. This advancement in quantitative fluorescence imaging holds significant value for tumor margin delineation and personalized therapy in precision oncology.

Graphical abstract

Highlights

• •HS-QFNet performs photosensitizer quantification with fused hyperspectral features
• •It is trained on optically diverse tissue phantoms to enhance model robustness
• •This approach outperforms traditional methods in phantom and in vivo studies
• •It supports tumor margin delineation and personalized photodynamic therapy

Biomedical discipline; Applied sciences; Machine learning

Introduction

Fluorescence imaging technology, with its advantages of high sensitivity and precise visualization, has become a key tool in modern clinical tumor diagnosis and treatment. The clinical efficacy of its core applications—fluorescence-guided surgery (FGS) and photodynamic therapy (PDT)—essentially relies on the selective accumulation of photosensitizers in target tissues.^1,2 In FGS applications, photosensitizers retained in pathological tissues generate fluorescence signals upon excitation, providing real-time visualization of lesion distribution that enhances surgical precision and improves patient outcomes.^3,4,5,6 The clinical success of 5-aminolevulinic acid (5-ALA)-induced protoporphyrin IX (PpIX) in brain tumor resection exemplifies this potential.^4,7 Similarly, PDT employs light-activated photosensitizers to selectively destroy diseased tissues,^8,9,10 achieving remarkable clinical success in the treatment of superficial tumors. Its precise photosensitizer localization enables targeted ablation while preserving adjacent healthy tissues.¹¹

Despite these advances, significant technical challenges impede broader clinical adoption. The inherent optical characteristics of biological tissues, notably absorption and scattering phenomena, introduce intricate nonlinear distortions in the transmission of fluorescence signals, thereby undermining the reliability of conventional fluorescence intensity-based quantification approaches.¹² In FGS, such signal attenuation or distortion directly compromises the accuracy of tumor boundary delineation, risking incomplete resection due to underdetection of marginal lesions or unnecessary healthy tissue removal due to false-positive signals.¹² For PDT, the significant interpatient variability in tissue optical properties, drug metabolism, and tumor microenvironment necessitates precise in vivo photosensitizer quantification to enable personalized dosimetry.^13,14 Through accurate photosensitizer quantification, clinicians can more clearly delineate tumor margins in FGS based on quantitative data, reducing subjective errors from intensity-only qualitative judgment and better distinguishing tumor from normal tissue, while dynamically adjusting treatment doses in PDT according to real-time acquired photosensitizer concentration data. Consequently, developing accurate fluorescence quantitative methods that correct tissue-induced signal distortion and provide near-real-time quantitative feedback is crucial for optimizing tumor margin identification in FGS and enabling precise drug delivery in PDT.^7,12

Current mainstream approaches for wide-field fluorescence quantification correction integrate hyperspectral imaging (HSI) technology with dual-band normalization.^15,16 HSI provides comprehensive spectral information across the entire field of view,¹⁷ while dual-band normalization compensates for optical interference by measuring diffuse reflectance at both excitation and emission bands.¹⁸ This integrated approach has demonstrated clinical utility in tumor margin detection and photosensitizer quantification.^15,16,19 Notable advancements include the spectrally constrained dual-band normalization algorithm proposed by Valdés et al.,^18,20 which demonstrated high accuracy in tissue phantom experiments using only two reflectance bands for fluorescence correction. Further innovations have enhanced sensitivity, quantitative accuracy, and contrast-to-noise ratio, as seen in the works of Jermyn et al.,²¹ Xie et al.,²² and Bravo et al.²³ Recent developments have also focused on improving calibration algorithms and imaging system designs to overcome limitations in visualization sensitivity and quantitative accuracy.^24,25,26

However, current dual-band calibration algorithms and their improved variants, despite achieving high spatial resolution and calibration accuracy, fundamentally rely on semi-empirical formulas. This reliance makes it challenging to effectively characterize the complex nonlinear coupling effects between fluorescence transmission and tissue absorption and scattering in biological tissues.²⁷ These methods are constrained by the assumptions of theoretical models and experimental conditions, often requiring tedious system-specific parameter adjustments when confronted with tissue heterogeneity or different imaging systems. This limits their precision and adaptability, hindering the accuracy of real-time FGS and precise quantification of photosensitizers for PDT. Thus, there is an urgent need for correction algorithms capable of overcoming empirical model restrictions and modeling intricate nonlinear relationships.

In recent years, deep learning has driven significant advancements in tissue optical characterization methodologies. Our team developed a multi-reference phantom-driven network (MR-Net) that uses deep convolutional architectures to accurately map optical attenuation coefficients from optical coherence tomography signals, significantly outperforming traditional methods.²⁸ Furthermore, deep learning technology provides crucial support for advancing intraoperative HSI, particularly in addressing challenges such as intraoperative tumor tissue identification and fluorescence quantification.²⁹ Innovative approaches integrating HSI with machine learning frameworks have demonstrated significant potential in real-time multi-tissue recognition and precise tumor segmentation.^30,31 Additionally, deep learning models based on a convolutional neural network (CNN) have achieved quantitative estimation of fluorescence component abundance and high-precision automated tumor classification,^32,33 which effectively mitigate tissue-induced optical distortions and set a strong benchmark for fluorescence-guided neurosurgical applications. Notably, Baumann et al.³⁴ introduced a deep learning-based method for intraoperative HSI calibration, enabling real-time calibration under dynamic operating room lighting conditions. These studies collectively underscore the transformative potential of deep learning in enhancing fluorescence quantification techniques.

Building upon this foundation, the present study introduces a fluorescence correction algorithm, Hyperspectral Quantitative Fluorescence Network (HS-QFNet), based on deep CNN and attention mechanisms. This approach aims to surmount the limitations of traditional algorithms by integrating features from wide-field hyperspectral fluorescence and diffuse reflectance images. HS-QFNet employs channel-spatial attention mechanisms to end-to-end learn the complex nonlinear mapping relationships between fluorescence signal distortion and tissue optical properties, ultimately predicting photosensitizer concentration distribution. The model was trained and validated on tissue-mimicking phantoms covering a broad range of biological optical properties and was directly applied to in vivo studies without fine-tuning. Experimental validation in phantom and animal models, with comparative assessments against traditional calibration algorithms, evaluates the quantitative performance advantages of HS-QFNet. This algorithm is anticipated to enable more accurate photosensitizer concentration prediction, providing a promising tool for personalized PDT treatment and FGS navigation, thereby enhancing the clinical potential of fluorescence imaging in precision oncology.

Results

Performance of the traditional dual-band calibration algorithm for photosensitizer concentration prediction in liquid phantoms

Figure 1 presents the hyperspectral characterization of PpIX solutions and tissue-simulating phantoms, demonstrating the performance of the dual-band calibration approach. The study encompassed three key components: (1) pure PpIX reference solutions (Figures 1A and 1B), (2) liquid phantoms with controlled scattering and absorption properties (Figures 1C and 1D), and (3) representative calibrated spectra (Figures 1E and 1F). For quantitative analysis, we selected a uniform 1,000 × 1,000 pixel central region of interest (ROI) from each phantom, using the average spectral intensity to represent fluorescence and diffuse reflectance characteristics at each wavelength. The calibration algorithm, employing nonlinear least squares fitting to determine empirical coefficient, showed good linear correlation (R² > 0.98) between photosensitizer concentration and emission band intensity (620–640 nm) in reference solutions (Figure 1B). Figures 1C and 1D depict the raw fluorescence and diffuse reflectance spectra of group A3,1 to A3,5 liquid phantoms, which were designed to exhibit a range of optical properties as detailed in Table 1. The dual-band algorithm demonstrated good performance: for the A3,2 phantom (with a true concentration of 2 μM), the calibration yielded a value of 2.06 μM (Figure 1E); similarly, for the A3,4 phantom (with a true concentration of 8 μM), the results indicated 8.72 μM (Figure 1F). Calculation revealed an average concentration calibration error of 4.6% across all phantoms. Sources of error include the intrinsic limitations of the semi-empirical algorithm, as well as potential factors such as inhomogeneity in liquid phantoms and system noise.

Figure 1.

Calibration results using the dual-band calibration algorithm

(A) Fluorescence intensity of pure PpIX solutions at different concentrations.

(B) Fitted calibration curve of PpIX fluorescence intensity versus concentration. Data are represented as mean (n = 3 independent measurements per concentration).

(C) Raw fluorescence spectra of A3,1 to A3,5 liquid phantoms.

(D) Raw diffuse reflectance spectra of A3,1 to A3,5 liquid phantoms.

(E) Calibration results for phantom A3,2.

(F) Calibration results for phantom A3,4.

Table 1.

Liquid phantom array

		Concentration
Phantom		Intralipid (w/v%)	Blood (v/v%)	PpIX (μM)
Group A	A1,1 to A1,5	0.25%	0.5%	1; 2; 4; 8; 16
	A2,1 to A2,5	0.5%	0.5%	1; 2; 4; 8; 16
	A3,1 to A3,5	1%	0.5%	1; 2; 4; 8; 16
	A4,1 to A4,5	2%	0.5%	1; 2; 4; 8; 16
Group B	B1,1 to B1,5	0.25%	1%	1; 2; 4; 8; 16
	B2,1 to B2,5	0.5%	1%	1; 2; 4; 8; 16
	B3,1 to B3,5	1%	1%	1; 2; 4; 8; 16
	B4,1 to B4,5	2%	1%	1; 2; 4; 8; 16

Validation of HS-QFNet’s accuracy in predicting photosensitizer concentration in liquid phantoms

Figure 2 compares the concentration of PpIX predicted by HS-QFNet with the concentration calibrated by the traditional dual-band calibration method mentioned earlier and shows the remapping results of the calibration concentration and error of the two methods in the liquid phantoms listed in Table 1. Specifically, Figures 2A and 2B show the color mapping of the mean absolute error (MAE) values of group A phantoms predicted by the traditional dual-band calibration method and HS-QFNet, respectively. The traditional method shows a relatively obvious red area, indicating a large error, while HS-QFNet exhibits lower error distribution. Figure 2C depicts the concentration distributions predicted by the two methods, where the ideal line represents a theoretically perfect prediction. The predictions of the traditional method exhibit a certain degree of deviation, while those of HS-QFNet are closer to the ideal line, and the 95% confidence interval corresponding to HS-QFNet is also significantly superior to that of the traditional method. Figure 2D presents the predicted results and standard deviations (SD) for 10 groups of phantoms, with each group consisting of 100 samples randomly drawn from the test set. The results demonstrate the excellent prediction stability and statistical reliability of HS-QFNet. Figure 2E shows the mean values and distributions of the predicted MAE for the two methods. For the proposed HS-QFNet, the MAE between the predicted and true concentrations is only 0.21 μM, with an SD of 0.16 for the mean value. In contrast, the traditional method yields an MAE of 0.65 μM and an SD of 0.54 μM. Notably, compared with the traditional dual-wavelength calibration method, the average error of HS-QFNet is reduced by 68%. These results indicate that HS-QFNet achieves higher accuracy in predicting photosensitizer concentration than the traditional dual-wavelength calibration method. The ablation experiments in Figure 2F demonstrate that the convolutional block attention module (CBAM) effectively improves the model’s prediction accuracy and stability, while the diffuse reflectance spectrum input provides key spectral band variations associated with tissue absorption and scattering. This allows the model to identify tissue absorption and scattering properties, implicitly disentangling their respective contributions to correct fluorescence signal distortion and enhance generalization ability.

Figure 2.

Photosensitizer concentration prediction of phantoms based on neural networks

(A) Predicted MAE map of traditional dual-band method.

(B) Predicted MAE map of HS-QFNet.

(C) Concentration prediction results of group A phantom by the traditional method and HS-QFNet. Shaded regions represent 95% confidence intervals.

(D) Leave-one-phantom-out (LOPO) cross-validation: MAE of phantoms in groups A1 and B4. Data are represented as mean ± SD (n = 100 randomly sampled pixels per phantom, 10 independent groups).

(E) Prediction average MAE and distribution of two methods. Data are represented as mean MAE ± SD (HS-QFNet: 0.21 ± 0.16 μM; traditional method: 0.65 ± 0.54 μM).

(F) Results of ablation experiments. Data are represented as MAE ±SD (without CBAM: 0.89 ± 0.54 μM; without diffuse reflectance: 0.21 ± 0.36 μM; HS-QFNet: 0.21 ± 0.16 μM).

Validation of HS-QFNet for in vivo photosensitizer concentration prediction in the mouse tumor model

Figure 3 effectively demonstrates the validation of HS-QFNet for in vivo prediction of photosensitizer concentration within mouse tumor models. In this study, nine mice bearing HeLa tumors were categorized into three groups, each receiving low, medium, and high concentrations of topical 5-ALA hexyl ester. Hyperspectral fluorescence and diffuse reflectance images were captured 4 h after administration. Figure 3A displays fluorescence intensity images of PpIX solutions at predetermined concentrations, facilitating the establishment of a concentration-intensity calibration curve with an R² value exceeding 0.98 (Figure 3B). Fluorescence intensity imaging of 20 μm cryosectioned tumor slices (Figure 3C) can be converted to PpIX concentrations using this calibration curve, providing in vivo true concentrations. As depicted in Figure 3D, the mean prediction error across all groups was 0.31 μM, accompanied by a mean squared error (MSE) of 0.146. However, the low-concentration group demonstrated notably lower accuracy, with an error of 0.55 μM, in contrast to the medium- and high-concentration groups, which exhibited errors of only 0.18 μM. This discrepancy is presumably ascribed to the combined effects of insufficient training data available for lower concentrations and the inadequate signal-to-noise ratio inherent to low-concentration samples. Figure 3E further showcases the dynamic monitoring of photosensitizer concentration changes in the mouse tumor during PDT treatment. The in vivo study underscores the considerable potential of HS-QFNet for accurate in vivo prediction of photosensitizer concentration.

Figure 3.

In vivo photosensitizer concentration prediction in mouse tumor models using HS-QFNet

(A) Fluorescence intensity of 20-μm-thick PpIX solutions at different concentrations.

(B) Fitted calibration curve of concentration versus fluorescence intensity.

(C) Fluorescence micrograph of mouse tumor section.

(D) Quantitative accuracy of neural network for in vivo photosensitizer concentration prediction.

(E) Monitoring of in vivo photosensitizer concentration dynamics during PDT treatment.

Discussion

This study presents an approach to address the challenge of accurate photosensitizer quantification in FGS and PDT. We developed a wide-field HSI system for the simultaneous acquisition of hyperspectral fluorescence and diffuse reflectance spectra (Figure 4). Our HS-QFNet algorithm represents a significant departure from traditional semi-empirical approaches by establishing an end-to-end deep learning framework that directly models the complex relationship between fluorescence signals and tissue optical properties (Figure 5). The integration of wide-field HSI with advanced convolutional neural networks and attention mechanisms has yielded a system capable of accurate quantification of photosensitizer concentration in shallow biological tissue environments. The training data for the model comes from liquid phantoms with a wide range of optical properties, ensuring its in vivo quantitative generalization. The experimental results demonstrate HS-QFNet’s superior performance across multiple dimensions. In controlled phantom studies, the algorithm achieved a remarkable 68% reduction in quantification error compared to traditional dual-band normalization methods, with MAE decreasing from 0.65 to just 0.21 μM. The in vivo validation using mouse tumor models further confirmed the clinical potential of our approach (Figure 6), showing strong correlation with true PpIX concentrations and reliable performance during near-real-time PDT monitoring.

Figure 4.

Schematics of the hyperspectral imaging system and fabrication of the phantoms

(A) Hyperspectral imaging system.

(B) Extraction of PpIX fluorescence and diffuse reflectance spectra.

(C) Constructed liquid phantom array for groups A and B. C_PpIX, protoporphyrin IX concentration.

Figure 5.

Schematic of the HS-QFNet network architecture

(A) Dataset construction.

(B) Feature extraction module and regression prediction module. Training and testing sets were derived from the prepared phantom data.

Figure 6.

In vivo prediction of photosensitizer concentration in mouse tumor models using HS-QFNet

(A) Workflow for network prediction and validation, consisting of three key steps: (1) in vivo prediction-acquisition of hyperspectral fluorescence and diffuse reflectance images from tumor models, followed by network-based concentration mapping; (2) solution calibration for establishing a reference standard; and (3) ex vivo validation through fluorescence imaging of tumor slices and calibration curve fitting to evaluate the quantitative accuracy of photosensitizer concentration.

(B) Spatial distribution map of predicted photosensitizer concentration within the tumor.

Traditional fluorescence quantification methods face fundamental limitations in clinical applications. The dual-band normalization algorithm proposed by Valdés et al.¹⁸ and its subsequent improvements^{21,22,23,24,25} rely on simplified physical assumptions (single-scattering approximations and homogeneous medium models) that cannot adequately characterize the complex optical properties of real tissues. While these methods have achieved progressive accuracy enhancements through various optimizations, including adaptive weighting coefficients,^22,23 enhanced calibration algorithms,²⁴ and improved hardware configurations,^21,25 they remain constrained by their empirical framework. Key limitations include (1) dependence on parameterized relationships requiring system-specific calibration, (2) inability to model nonlinear fluorescence-tissue interactions, and (3) computational complexity hindering real-time application. Recent work by Black et al.³² proposed deep-learning-based hyperspectral correction and unmixing frameworks that explicitly incorporate physical models and known endmember spectra to improve fluorescence-guided brain tumor surgery. Their approach focuses on attenuation correction and spectral unmixing to estimate endmember abundances, with particular emphasis on generalizability across patients and tissue types. Our work, by contrast, offers a complementary solution tailored to superficial tumor scenarios, leveraging end-to-end learning to bypass the need for predefined endmembers or complex system calibration. Unlike deep spectral unmixing methods that focus on separating the abundance of known spectral components, HS-QFNet aims to correct tissue optical distortion and establish the mapping from distorted spectra to real concentration, which does not depend on pure spectral prior information and is more consistent with the clinical needs of rapid quantitative detection.

The HS-QFNet algorithm represents a paradigm shift by employing deep convolutional networks to directly learn the nonlinear mapping between spectral distortions and tissue optical properties. Our architecture combines three key components: (1) separate initial feature extraction branches for fluorescence spectra and diffuse reflectance spectra, coupled with late-stage feature fusion, (2) CBAM attention mechanisms for dynamic feature weighting, and (3) an end-to-end learning framework that eliminates empirical assumptions. This integrated approach enables precise identification of characteristic spectral features (PpIX peaks and hemoglobin valleys) while modeling complex fluorescence-reflectance interactions. The CNN’s hierarchical structure and attention mechanisms (channel attention for fluorescence-reflectance correlations; spatial attention for critical spectral bands) work synergistically to overcome the oversimplifications inherent in traditional methods. Phantom validation demonstrated exceptional performance, with 68% error reduction compared to conventional approaches. It should be noted that HS-QFNet is not a photosensitizer-specific model but a generalizable framework. For organic photosensitizers with similar spectra to PpIX, transfer learning enables fast fine-tuning; for inorganic photosensitizers with large spectral shifts, full retraining on new datasets is recommended to ensure accuracy.

The system’s HSI hardware provides the essential foundation for accurate quantification. Utilizing a liquid crystal tunable filter (LCTF), the system acquires wide-field, continuous spectral data (420–720 nm) with high resolution and rapid wavelength switching. This configuration offers three key advantages: (1) complete spectral-spatial information capture per pixel, (2) elimination of sampling limitations inherent in point-detection systems, and (3) fast imaging capability for clinical applications. The tight integration of this advanced hardware platform with the HS-QFNet algorithm creates a promising solution for in vivo photosensitizer quantification, with potential to address critical needs in both FGS tumor delineation and PDT dose personalization.

In conclusion, this study proposes HS-QFNet, a fluorescence quantification correction algorithm based on CNN, to address the critical clinical need for personalized FGS navigation and PDT. By fusing features from wide-field hyperspectral fluorescence and diffuse reflectance images of phantoms, and incorporating the CBAM mechanism, HS-QFNet establishes an end-to-end non-linear mapping model between fluorescence signals and tissue optical properties, enabling high-precision spatial concentration mapping of PpIX in both tissue-simulating phantoms and in vivo settings. Experimental validation demonstrates significant improvements in quantification accuracy compared to traditional methods and excellent generalization capability, providing a powerful tool for improving tumor boundary identification accuracy in FGS and supporting personalized photosensitizer dose control for PDT.

Limitations of the study

While this study demonstrates the potential of HS-QFNet, several limitations warrant consideration. First, regarding model design: because the phantom’s optical properties span the reported range of cervical cancer tissues, the parameter space remains insufficiently broad; the current phantoms do not cover highly vascularized tissues or those with unique optical properties, and applying the model to tissues with optical properties outside this range may therefore introduce significant errors. Such tissues require expanded training datasets to achieve reliable quantification. Moreover, the phantom primarily represents a homogeneous liquid environment and cannot fully replicate the multilayered architecture and structural heterogeneity of real biological tissues. This simplification restricts the model’s applicability to single-layer, superficial tissues and limits its broader translational relevance. Second, with respect to experimental validation: although the animal study provides adequate data for preliminary verification, its relatively small sample size may not be able to fully capture biological variability that could impact model performance in extended applications. Increasing the sample size would strengthen the robustness and generalizability of the findings. Third, unresolved intraoperative factors impede clinical translation: the study does not account for the complexities of real surgical settings, including tissue motion, ambient light interference, and the stringent requirements for real-time processing—latencies in data handling could compromise intraoperative decision-making. Addressing these challenges is essential for successful clinical adoption.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Defu Chen (defu@bit.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data reported in this paper will be shared by the lead contact upon request.
• •This paper reports original custom code. The code is available from the lead contact upon request.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Key Research and Development Program of China (2023YFB3609100), National Natural Science Foundation of China (62205025 and 62227823), Beijing Municipal Natural Science Foundation (7222309), and Beijing Institute of Technology Research Fund Program for Young Scholars (XSQD-202123001).

Author contributions

Conceptualization, D.C. and S. Hao; methodology, D.C., S. Hao, and Xinpeng Zhang; investigation, S. Hao, Xinpeng Zhang, S. Han, and Y.X.; data curation, S. Hao, Xinpeng Zhang, S. Han, and Y.X.; formal analysis, S. Hao and Xiwan Zhang; hardware development, S. Hao and S. Han; software, S. Hao and Xinpeng Zhang; validation, S. Hao and Xinpeng Zhang; visualization, S. Hao and Xinpeng Zhang; writing – original draft, S. Hao; writing – review and editing, D.C., S. Hao, Xinpeng Zhang, Xiwan Zhang, and Y.X.; funding acquisition, H.Q., Y.G., and D.C.; project administration, D.C.; supervision, D.C., H.Q., and Y.G.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Bovine whole blood	Beijing Borxi Technology Co., Ltd.	Cat No.: BWB-100
Chemicals, peptides, and recombinant proteins
Protoporphyrin IX (PpIX) disodium salt	Sigma-Aldrich	CAS: 140898-91-5
Dimethyl sulfoxide (DMSO)	Beyotime Biotech Inc	ST038
Phosphate-buffered saline (PBS)	Cytiva	Cat No.: SH30256.01
Intralipid 20%	CSL Behring Pharma GmbH	NMPA Approval No.: H19993197
5-ALA hexyl ester-gel mixture (Hexvix)	Shanghai PharmPrep	CAS: 106-60-5
Experimental models: Cell lines
HeLa cells	ATCC	N/A
Experimental models: Organisms/strains
Female BALB/c nude mice	SPF (Beijing) Biotechnology Co., Ltd.	N/A
Software and algorithms
HS-QFNet source code	This paper	Available from Lead Contact
Python	3.11	https://www.python.org/
PyTorch	2.2.2	https://pytorch.org/
Other
405 nm laser	Changchun New Industries Optoelectronics Technology Co., Ltd.	MDL-III-405-150mW
High-power white LED ring light guide	Olympus Corporation	SZ2-LGR
Stereomicroscope	Olympus Corporation	SZ61TR
Liquid Crystal Tunable Filter (LCTF)	Cambridge Research and Instrumentation	VariSpec
CMOS camera	Tucsen Photonics	FL-20BW
Inverted microscope	Olympus Corporation	IX70

Experimental model and study participant details

Female BALB/c nude mice (4–6 weeks old) were purchased from Sipeifu (Beijing, China) and housed under standard specific-pathogen-free (SPF) conditions with controlled light-dark cycles, temperature, and humidity. All animals had free access to food and water. Ethics approval for the experiments reported in the submitted manuscript on animal experimentation was granted, and all animal experiments were conducted in accordance with the procedures approved by the Institutional Animal Ethics Committee of Beijing Institute of Technology (Approval No. BIT-EC-SCXK-2019-0010-M-046). The detailed information of the phantom and other materials is specified in the main text and method details.

Method details

Wide-field spectrally resolved quantitative fluorescence imaging system

To enable precise quantitative analysis of photosensitizers, we developed a custom wide-field hyperspectral imaging system, as illustrated in Figure 4A. This system consists of two main components: excitation light sources and a hyperspectral imaging module. For fluorescence excitation, a 405 nm laser (MDL-III-405-150mW, CNI, China) is employed, while a high-power white LED ring light guide (SZ2-LGR) serves as the source for diffuse reflectance spectroscopy. The laser output is delivered via an optical fiber, ensuring efficient coupling. The optical signals collected by the stereomicroscope (SZ61TR, Olympus Corporation, Japan) objective are spectrally resolved using a continuously liquid crystal tunable filter (LCTF, VariSpec, Cambridge Research and Instrumentation, USA). To minimize image shift and aberrations caused by non-parallel light incidence into the LCTF, a 4f relay system was employed to ensure parallel light incident, thereby optimizing imaging performance. The final optical signals are captured by a high-sensitivity complementary metal-oxide-semiconductor (CMOS) camera (FL-20BW, Tucsen Photonics, China), which is specifically designed for low-light imaging and long-exposure applications.

The LCTF operates within the 420–720 nm range, offering a full width at half maximum (FWHM) of approximately 10 nm, peak transmittance >30%, and a rapid wavelength switching speed of 50 ms, enabling high-precision and fast spectral resolution. The CMOS camera features a quantum efficiency >65% in the 400–650 nm range, a 5472 × 3648 pixel sensor, and provides 16-bit digital output.

Data acquisition

As shown in Figure 4B, the system acquires data with the LCTF scanning from 420 to 720 nm in 10 nm increments (400 ms switching duration) while the CMOS camera operates at 200 ms exposure. In fluorescence mode, samples were excited by the 405 nm laser (10 mW/cm² surface irradiance) through a collimator, yielding 31 fluorescence spectral images. In diffuse reflection mode, the system utilizes the white LED ring light guide to collect an equivalent 31 diffuse reflection spectral images dataset. Raw hyperspectral data cubes (5472 × 3648 × 31) undergo preprocessing including dark current correction (averaged dark-field images subtraction) and flat-field correction (using SRT-99-120 standard diffuse reflector, Labsphere), followed by interpolation-based downsampling to enhance processing efficiency. All experiments were conducted under darkroom conditions to eliminate ambient light interference.

Liquid phantom construction

To simulate fluorescence propagation in biological tissues, we fabricated a series of tissue-mimicking solutions with varying optical properties. Protoporphyrin IX (PpIX) disodium salt (Sigma) was initially dissolved in dimethyl sulfoxide (DMSO) to prepare the PpIX stock solution, which was then diluted into phosphate-buffered saline (PBS) as the base matrix. To control scattering properties, Intralipid 20% was added at different concentrations (0.25%, 0.5%, 1%, and 2% intralipid), while bovine whole blood was incorporated at different concentrations (0.5% and 1% blood) to modulate absorption characteristics. The phantom groups (A and B) with specific scattering/absorption combinations correspond to the constructed phantom arrays shown in groups A and B in Figure 4C, respectively. These formulations produced tissue-simulating phantoms with precisely controlled absorption and scattering properties, maintaining accurate representation of the photosensitizer’s true concentration.³⁵ During preparation and measurement, all liquid phantoms were prepared as a single layer to minimize ambiguity in inverse estimation caused by the superposition of fluorescence signals from multiple depths. Recognizing the critical role of surfactants in stabilizing aqueous PpIX solutions, we incorporated 0.1% Tween 20 in all phantom formulations to prevent aggregation-induced fluorescence quenching.³⁶ Each phantom was thoroughly vortexed to ensure homogeneity.

The absorption and scattering coefficient of each phantom were characterized using a calibrated integrating sphere system. The results showed that at 405 nm, a wavelength commonly used in fluorescence imaging, the scattering coefficient (μ_s) of the phantoms effectively spanned 5-25 mm⁻¹, while the absorption coefficient (μ_a) was accurately confined to 0.25–0.5 mm⁻¹. This set of parameters is highly relevant for matching biological tissues: it not only encompasses the optical characteristics of typical thin biological layers such as the skin epidermis and mucosal tissues, but also spans the full reported range of optical parameters for cervical epithelial tissues (including both normal and pathological states)—the primary focus of this study. Consequently, the phantoms can faithfully reproduce the light-propagation behavior observed in these target tissues.

These precisely characterized optical properties, which are highly matched to shallow biological tissues (especially cervical epithelium), were directly used to construct the training dataset for the HS-QFNet. In order to eliminate interference from fluorescence signals at different depths, the thickness of the liquid phantom is strictly controlled during the measurement process. This design ensures that the model derived from the phantoms can be directly transferred to the fluorescence analysis of in vivo biological tissues without the need for additional system calibration. It not only reduces errors caused by differences in optical environments between in vitro and in vivo settings but also lays a crucial foundation for the model to achieve generalization across devices and clinical scenarios, providing reliable experimental support for subsequent studies on precise fluorescence diagnosis of cervical diseases.

Traditional dual-band fluorescence calibration

As the baseline correction methods, we implemented hyperspectral imaging with dual-band normalization. This algorithm compensates for tissue-induced fluorescence distortion by utilizing diffuse reflectance characteristics at both excitation and emission bands. The corrected fluorescence spectrum F_(i,j)corr(λ) was calculated using the dual-band normalization:

$$F_{\left(i,j\right)corr}\left(\lambda \right)=\frac{F_{\left(i,j\right)}\left(\lambda \right)}{{\left({\Phi }_{ex}\right)}^a\times {\left({\Phi }_{em}\right)}^b}$$ (Equation 1)

Where F_(i,j)(λ) is the raw measured fluorescence spectrum. Φ_ex and Φ_em represent the integral value of diffuse reflectance over the fluorescence excitation and emission bands respectively, with empirical coefficients a and b determining through experimental fitting. Given the linear relationship between photosensitizer concentration and fluorescence intensity within a certain range, the fluorescence intensity per unit concentration is measured, and the following formula is fitted:

$$\mathit{arg}\underset{C}{\mathit{min}}{\Vert f_{basic}\left(\lambda \right)·C-F_{corr}\left(\lambda \right)\Vert }_2^2$$ (Equation 2)

Here, f_basic(λ) represents the true fluorescence intensity per unit concentration of the photosensitizer, F_corr(λ) is the corrected fluorescence intensity, and C is the fitted photosensitizer concentration. Pixel-wise application of this method yields comprehensive quantitative concentration maps.

Neural network-based photosensitizer concentration quantification algorithm

To overcome the limitations of empirical models, we developed HS-QFNet, a CNN-based algorithm for photosensitizer quantification. The model was developed and validated using a comprehensive dataset of PpIX liquid phantoms with precisely controlled concentrations and optical properties. To enhance generalizability, we implemented a robust preprocessing pipeline comprising: (1) fluorescence intensity normalization to a standard scale, (2) diffuse reflectance spectra standardization, and (3) controlled noise injection to simulate measurement variability. As shown in Figure 5A, paired fluorescence and diffuse reflectance spectra were extracted from hyperspectral images for each spatial location, forming dual-channel 1D sequences (31 × 2) that served as network inputs, with corresponding true PpIX concentrations as regression targets.

The HS-QFNet architecture (Figure 5B) processes these spectral inputs through a hierarchical feature extraction module, which includes independent initial feature extraction branches for two channels (fluorescence spectrum and diffuse reflectance spectrum), each branch containing a continuous convolutional layer (3 × 1 kernel, ReLU activation function, batch normalization) and a max pooling layer; After feature fusion, three residual blocks and a convolutional block attention module (CBAM) for channel spatial feature optimization are used to perform channel attention optimization first, followed by spatial attention optimization; The extracted features are then input into a regression head composed of a three-layer multilayer perceptron (MLP), which gradually reduces the dimensionality while converting the enhanced spectral representation into quantitative concentration prediction.

The CBAM attention module was specifically selected for this hyperspectral fluorescence quantification task due to its dual advantages of targeted feature enhancement and lightweight design. Its channel attention mechanism adaptively adjusts the weight distribution between fluorescence and diffuse reflectance spectral channels, prioritizing features related to photosensitizer concentration and tissue optical properties. Meanwhile, the spatial attention component focuses on key spectral regions in the lesion area, suppressing background interference to strengthen effective signal recognition. Importantly, CBAM introduces minimal additional computational overhead, ensuring the model maintains near-real-time inference performance for high-dimensional hyperspectral data.

The model training process in this study is designed as follows: The validation phantoms are completely independent of the training phantoms, and the dataset is split into a training set and a test set at an 80/20 ratio, with the division performed at the level of entire phantoms (i.e., all pixels from a single phantom were assigned to either the training or test set, not split between them). This ensures the training and test sets are derived from fully distinct phantoms with no overlap. The total number of training epochs is set to 1000, with a batch size of 1024. The Adam optimizer is adopted for training, combined with the ReduceLROnPlateau learning rate scheduling strategy to dynamically adjust the training rhythm. To improve the model’s generalization ability, Leave-One-Phantom-Out (LOPO) cross-validation is employed: in each iteration, all samples of one phantom are excluded as an independent test set, while the samples of all other phantoms form the training set. Finally, the optimal model is selected using the minimum loss value on the independent test set as the core evaluation metric, with the evaluation process completely isolated from the training data.

To verify the roles of CBAM and diffuse reflectance spectroscopy, ablation experiments were conducted. Two ablated models were trained under identical configurations: one without CBAM and the other without diffuse reflectance data. Performance comparisons between these variants and the full model quantify the specific contributions of each component.

In vivo animal experiment

To validate the phantom-trained HS-QFNet’s performance under in vivo conditions, we established a cervical cancer model by subcutaneously inoculating HeLa cells in female BALB/c nude mice. Tumor implantation was specifically performed in superficial dorsal locations to facilitate subsequent optical measurements. The experimental protocol involved topical application of 5-ALA hexyl ester-gel mixture (Hexvix, Shanghai PharmPrep, China) to the tumor region, followed by a 4-h metabolic period to allow for PpIX accumulation through endogenous conversion.³⁷ Subsequently, hyperspectral fluorescence and diffuse reflectance imaging were performed to predict the intratumoral photosensitizer concentrations. For precise quantification, mice were euthanized immediately after imaging for tumor extraction. The excised tumors were embedded in optimal cutting temperature (OCT) compound and cryosectioned into 20 μm slices.

In the visible spectral range employed in this study, the scattering mean free path of biological soft tissues is typically on the order of 50–100 μm, indicating that photons undergo significant scattering events only after propagating over distances comparable to or larger than this scale. In our experiment, the thickness of tumor tissue sections was strictly controlled to 20 μm, well below the scattering mean free path, resulting in a substantial reduction in the probability of photon scattering. Under this condition, light propagation is dominated by ballistic or quasi-ballistic photons, and scattering-induced distortions become negligible. Furthermore, absorption effects in such ultra-thin sections are minimized and exhibit approximate linearity. For a short optical path length L = 20 μm, the Beer–Lambert attenuation term $e^{{-\mathrm{\mu }}_{\mathrm{a}}L}$ indicates that the absorption-induced light loss is negligible for concentration-dependent fluorescence quantification. This minimal absorption ensures that the fluorescence intensity detected from the ultra-thin sections is minimally attenuated, maintaining a stable and approximately linear relationship with the local photosensitizer concentration, and our experiment has verified this through validation.

Experiments were performed using an Olympus IX70 inverted microscope for fluorescence microscopic observation under excitation with a 405 nm laser. Mice treated only with gel were used as controls for autofluorescence correction. The experiments combined the rapid cryofixation technique for OCT compounds, with imaging measurements conducted immediately after sectioning. Meanwhile, analysis was performed by averaging the measured values from ≥5 non-overlapping regions in each tissue layer, which effectively reduces the variability caused by spatial heterogeneity. For quantitative validation, we established a thickness-matched calibration curve using PpIX solutions of known concentrations confined between glass slides with 20 μm spacers. Fluorescence intensity measurements for each standard solution were repeated independently ≥3 times to account for experimental variability, and the mean value (± standard deviation, SD) was used for curve fitting to generate a concentration-fluorescence relationship. The fluorescence intensities of the tumor sections were then converted to true PpIX concentrations via the calibrated curve, allowing for a direct comparison with network predictions to evaluate quantitative accuracy.

Figure 6A illustrates the workflow for network prediction and validation, which comprises three critical steps: 1. In vivo prediction-acquisition of hyperspectral fluorescence and diffuse reflectance images from tumor models, followed by network-based concentration mapping; 2. Solution calibration-establishing a reference standard; and 3. Ex vivo validation-fluorescence imaging of tumor slices and calibration curve fitting to ascertain the true photosensitizer concentration. As depicted in Figure 6B, hyperspectral fluorescence and diffuse reflectance imaging were conducted to predict the intratumoral photosensitizer concentration.

To explore potential clinical utility, we implemented near-real-time photosensitizer monitoring during PDT. Following 4-hour 5-ALA hexyl ester application, tumors were irradiated using a 620 nm organic light-emitting diode (OLED) panel (100 mW/cm²) with concurrent hyperspectral imaging at 10-min intervals. This enabled network-based tracking of dynamic photosensitizer concentration changes throughout the therapeutic intervention. All animal experiments were conducted in accordance with the procedures approved by the Institutional Animal Ethics Committee of Beijing Institute of Technology.

Quantification and statistical analysis

Quantification, analysis, and visualization of results from the model simulations were carried out using a combination of Python and Microsoft Excel. Details of the methods used to aggregate data represented in all figures are specified in the method details section or in the figure captions.