Side-viewing probe for lesion depth mapping on the left ventricle epicardium with near-infrared spectroscopy Authors: Danyang Cheng, Jonah A. Majumder, Haiqiu Yang, Deepak Saluja, Kenneth R. Laurita, Andrew M. Rollins, Christine P. Hendon Published: 2026 ABSTRACT Epicardial radiofrequency ablation can fail when lesions are not sufficiently deep or transmural, yet intraoperative feedback remains largely indirect. This study presents a fiber-based, side-viewing near-infrared spectroscopy (NIRS) probe with multiple source-detector separations (SDS) to enable depth-sensitive mapping of lesions on the porcine left-ventricular epicardium. Monte Carlo simulations predicted progressively deeper sampling with increasing SDS, motivating the use of multi-separation acquisition for depth-resolved contrast. Experiments were performed on 11 porcine hearts with 133 irrigated epicardial lesions spanning a wide depth range, with lesion depth ground truth reconstructed from post-stain gross section measurements. SDS-dependent spectral signatures were observed across lesions with depths greater than 4 mm, lesions with depths ≤ 4 mm, adipose tissue, and untreated epicardial muscle, and optical indices capturing these patterns were identified for lesion and adipose classification, as well as for lesion-depth sensitivity. Lesion and adipose indices achieved strong receiver operating characteristic (ROC) performance across SDS (lesion AUC 0.87–0.91; adipose AUC 0.94–0.97), and depth-sensitive indices exhibited monotonic trends with lesion depth (R² up to 0.97). Applying a random forest lesion mask enabled depth-sensitive maps that were consistent with variations in the ground truth.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-50096-y. Introduction Radiofrequency ablation (RFA) is a cornerstone therapy for cardiac arrhythmias, where long-term success depends on creating lesions that are sufficiently deep, contiguous, and, if necessary, transmural to permanently disrupt pathological conduction1,2. In ventricular applications, arrhythmogenic substrates that are targets for ablation can extend deep within the left ventricular wall, and epicardial approaches are often required3,4. Therefore, the ability to achieve and verify adequate lesion depth is particularly crucial for effective ablation therapy5,6. However, confirming lesion extent during the procedure remains challenging7. Effective lesion formation is typically inferred from indirect surrogates, such as delivered power, ablation duration, catheter temperature, impedance changes, and contact force8,9. These measures correlate with average lesion size but do not directly quantify tissue necrosis at the ablation site10. Moreover, epicardial mapping and ablations frequently encounters adipose tissue which can insulate myocardium, dampen signals, and reduce the efficacy of radiofrequency energy delivery4,11. From an optical standpoint, adipose is lipid-rich and exhibits distinct near-infrared signatures that can confound lesion identification if not explicitly accounted for. These limitations underscore the need for intraoperative tools to accurately confirm lesion formation in real-time while remaining robust to tissue heterogeneity, including adipose, thereby improving the efficacy and safety of epicardial ablation therapy. Conventional ablation workflows rely on electroanatomic mapping (EAM) to reconstruct chamber geometry, guide catheter navigation, and lesion creation7,12. However, EAM-derived electrical measurements cannot indicate lesion depth directly. Intracardiac imaging modalities, such as magnetic resonance imaging (MRI) and computed tomography (CT), can visualize anatomy, fibrosis, and scars, but routine real-time lesion assessment remains challenging13–15. High-resolution optical imaging techniques, such as optical coherence tomography (OCT) and polarization-sensitive OCT, can resolve microstructural features and have been explored for assessing ablation-related tissue changes16–18. However, OCT’s shallow penetration (~ 1–2 mm) limited its ability to assess deeper tissue and lesion transmurality19–21. This limitation is important in ventricular applications, where the wall is thick. Ultraviolet autofluorescence hyperspectral imaging can delineate ablation lesions on the tissue surface, but validated quantitative lesion depth or transmurality assessment in a real-time percutaneous setting remains limited22,23. Near-infrared spectroscopy (NIRS) has emerged as a useful optical technique for probing tissue composition and structure in a depth-sensitive manner across various biomedical applications24–26. By analyzing broadband reflected light with appropriate models or data-driven approaches, these methods can estimate information about absorption, scattering, and layer thickness, enabling depth-resolved sensing in layered tissues and quantitative estimation of top-layer thickness on the order of sub-millimeter27. Optical engineering advances have also produced compact side-viewing probes that deliver and collect light laterally over a broad wavelength range, supporting multimodal imaging and spectroscopy within narrow lumens, albeit often at the cost of complex freeform micro-optics and increased distal-tip size28. Building on this foundation, our group has previously developed forward-viewing single-detector NIRS catheters and compact spectroscopic systems that acquire depth-sensitive spectra from cardiac tissue, enabling both lesion monitoring and substrate characterization29–31. Using these instruments, near-infrared measurements have been used to predict irrigated lesion depth and width, track lesion growth during energy delivery, and identify adipose regions and coronary vessels in atrial and ventricular myocardium, including high-density left atrial mapping and epicardial mapping on human ventricles32,33. In these prior studies, the optical geometry was predominantly forward-viewing, with a single detector adjacent to the source fiber. These fibers were integrated into standard 7 Fr RFA catheters, which restricted the achievable source-detector separation (SDS) range to smaller than 2.3 mm and reduced sensitivity to deep lesions and catheter tissue contact greater than 60 degrees34,35. A side-viewing multi-detector design overcomes these constraints by enabling larger and more varied SDS values. This allows the probe to sample multiple depth bands within the tissue, with smaller SDS emphasizing more superficial contrast and larger SDS providing greater sensitivity to subsurface tissue changes, thereby offering complementary information for lesion depth estimation. In addition, a side-viewing, fiber-based architecture can be packaged into a slim, elongated form factor that is compatible with the contemporary electroanatomic mapping catheters, including grid36 and ray-shaped37,38 designs, increasing the achievable SDS to more than 9 mm, thereby laying the groundwork for realistic clinical integration of depth-resolved NIRS lesion assessment. In this work, we introduce a fiber-based, side-viewing NIRS mapping approach designed to provide depth-sensitive assessment of radiofrequency lesions on the left ventricular epicardium. Through a combination of simulation, benchtop validation, and ex vivo experiments, we demonstrate that multi-detector NIRS can both identify ablated tissue and predict lesion depth, supporting progress toward real-time epicardial lesion assessment. Results Spectral signatures Figure 1 summarizes the spectral signatures acquired with various SDS values from the pig left ventricular epicardium across lesions with depths greater than 4 mm, lesions with depths ≤ 4 mm, adipose, and untreated epicardial muscle. The median lesion depth within the study was 4 mm, and was used as a threshold for deep and shallow lesions for spectral visualization and optical index development rather than a physiologically defined boundary. Spectra were grouped by tissue class, and the mean relative spectra with 95% confidence intervals were plotted to visualize class-dependent differences across the measured wavelength range within each SDS. At smaller SDS of 1 and 2 mm, the relative reflectance in the 820–1100 nm range increased with lesion severity, with lesions > 4 mm exhibiting higher reflectance than lesions ≤ 4 mm and lesions ≤ 4 mm higher than untreated myocardium. In contrast, adipose showed greater reflectance relative to the other classes in the 500–620 nm range. At larger SDS of 4 and 6 mm, spectra of lesions ≤ 4 mm closely overlapped with untreated epicardial muscle spectra across the measured band, whereas lesions > 4 mm exhibited a distinct spectral shape with reduced reflectance in the 650–850 nm range and increased reflectance in the 870–1100 nm range compared with both muscle and lesions ≤ 4 mm. The derivative spectra further emphasized adipose-specific differences, most prominently in the 580–620 nm and 980–1070 nm bands. To further compare tissue-class separation across SDS, Supplementary Fig. 1 presents the selected SDS-specific optical-index values for lesions > 4 mm, lesions ≤ 4 mm, adipose tissue, and untreated epicardial muscle, along with their statistical comparisons. Unless otherwise noted, all figures in the Results section are derived from ex vivo porcine experiments. Fig. 1SDS-dependent spectral morphologies of ex vivo porcine left ventricular epicardium. Normalized relative reflectance spectra with their derivatives are shown for SDS = 1 mm (A, E), 2 mm (B, F), 4 mm (C, G), and 6 mm (D, H). A - D: Mean spectra (solid) with 95% confidence intervals (shaded) for lesions of depth > 4 mm, lesions ≤ 4 mm, where 4 mm was used as a practical cutoff for visualization and OI development, together with adipose and untreated epicardial muscle; E - H: The corresponding first-order derivative spectra. SDS-dependent spectral morphologies of ex vivo porcine left ventricular epicardium. Normalized relative reflectance spectra with their derivatives are shown for SDS = 1 mm (A, E), 2 mm (B, F), 4 mm (C, G), and 6 mm (D, H). A - D: Mean spectra (solid) with 95% confidence intervals (shaded) for lesions of depth > 4 mm, lesions ≤ 4 mm, where 4 mm was used as a practical cutoff for visualization and OI development, together with adipose and untreated epicardial muscle; E - H: The corresponding first-order derivative spectra. Figure 2 illustrates that both the lesion optical index (LOI) and adipose contrast index (ACI) families demonstrated strong class separation across the tested range of SDS. Lesion identification based on the LOIs yielded consistently high receiver operating characteristic (ROC) performance, with area under the curve (AUC) values in the range of 0.87–0.91, and the highest AUC was observed at SDS = 2 mm at 0.913. Adipose identification using ACIs also performed strongly, producing near-ceiling ROC curves with AUC values between 0.94 and 0.97, peaking at SDS = 2 mm. While a modest variation with SDS was observed, the overall performance remained robust across 1–6 mm. Fig. 2Ex vivo ROC-based evaluation of LOIs and ACIs across SDS. ROC curves and corresponding AUC values are shown for lesion versus non-lesion differentiation using the LOI at SDS = 1 mm (A), 2 mm (B), 4 mm (C), and 6 mm (D), and for adipose versus non-adipose differentiation using the ACI at SDS = 1 mm (E), 2 mm (F), 4 mm (G), and 6 mm (H). Ex vivo ROC-based evaluation of LOIs and ACIs across SDS. ROC curves and corresponding AUC values are shown for lesion versus non-lesion differentiation using the LOI at SDS = 1 mm (A), 2 mm (B), 4 mm (C), and 6 mm (D), and for adipose versus non-adipose differentiation using the ACI at SDS = 1 mm (E), 2 mm (F), 4 mm (G), and 6 mm (H). Table 1 showed that, for classifying spectra acquired over lesions versus non-lesion epicardium, the leave-one-heart-out random forest models achieved their best overall performance at the smaller SDS. Accuracy was 0.854 and 0.853 at SDS = 1 and 2 mm, respectively, and decreased to 0.802 at 4 mm and 0.771 at 6 mm. Sensitivity followed the same pattern, dropping from 0.749 to 0.742 at 1 and 2 mm to 0.699 at 4 mm and 0.688 at 6 mm, whereas specificity remained comparatively high across separations at 0.916, 0.924, 0.873, and 0.840. Table 1Leave-one-heart-out random forest performance for classifying lesion versus non-lesion spectra across SDS. Mean (± standard deviation) accuracy, sensitivity, and specificity are reported for models trained and evaluated separately at SDS = 1, 2, 4, and 6 mm using the optical-index feature set.SDS (mm)AccuracySensitivitySpecificity10.854 ± 0.0270.749 ± 0.1260.916 ± 0.07420.853 ± 0.0250.742 ± 0.1460.924 ± 0.08740.802 ± 0.0730.699 ± 0.1820.873 ± 0.15060.771 ± 0.0860.688 ± 0.2460.840 ± 0.171 Table 1Leave-one-heart-out random forest performance for classifying lesion versus non-lesion spectra across SDS. Mean (± standard deviation) accuracy, sensitivity, and specificity are reported for models trained and evaluated separately at SDS = 1, 2, 4, and 6 mm using the optical-index feature set.SDS (mm)AccuracySensitivitySpecificity10.854 ± 0.0270.749 ± 0.1260.916 ± 0.07420.853 ± 0.0250.742 ± 0.1460.924 ± 0.08740.802 ± 0.0730.699 ± 0.1820.873 ± 0.15060.771 ± 0.0860.688 ± 0.2460.840 ± 0.171 Leave-one-heart-out random forest performance for classifying lesion versus non-lesion spectra across SDS. Mean (± standard deviation) accuracy, sensitivity, and specificity are reported for models trained and evaluated separately at SDS = 1, 2, 4, and 6 mm using the optical-index feature set. Depth and transmurality sensitivity of optical indices Figure 3 shows how depth-sensitive indices vary systematically with lesion severity when lesions are grouped by measured depth or by transmurality. Here, transmurality was quantified using the lesion ratio, defined as lesion depth divided by the full ventricular wall thickness. Across all SDS, the selected indices exhibited monotonic trends with lesion depth, and the linear fits to the bin means showed strong agreement with depth, with coefficient of determination (R2) values of 0.90, 0.97, 0.97, and 0.93 for SDS values of 1, 2, 4, and 6 mm, respectively. The slope direction depended on how each selected optical index was defined. The indices decreased with increasing depth (β = −3.40 and − 0.0229) when SDS = 1 and 4 mm, whereas the indices increased (β = 2.46 and 0.0281) when SDS = 2 and 6 mm. Multiple-comparison testing further supported depth sensitivity by showing statistically significant differences across most depth-bin pairs, while “ns” indicates specific non-significant comparisons, typically concentrated at the shallowest and deepest depth bins where variability increased. When evaluating against the lesion ratio, the indices again demonstrated approximately monotonic behavior, with the linear fits yielding R2 values of 0.94, 0.97, 0.89, and 0.72 for SDS values of 1, 2, 4, and 6 mm, respectively. The slope followed the same directionality as the depth-binned analysis, with negative trends at SDS = 1 mm and 4 mm (β = −35.16 and − 0.2438) and positive trends at SDS = 2 mm and 6 mm (β = 29.28 and 0.2289). The reduced R2 at SDS = 6 mm suggests that the binned means deviate more from a linear trend, with points showing a less consistent monotonic progression across transmurality bins. Fig. 3Ex vivo depth- and transmurality-sensitive behavior of the selected optical indices. A-D: Index means across 1-mm lesion-depth bins for SDS = 1, 2, 4, and 6 mm. E-H: The same indices across 10% lesion-ratio bins as an indicator for transmurality. Red points represent bin-averaged index values, with variability shown as the mean ± standard deviation (SD) (horizontal and vertical error bars). The blue lines indicate linear fits to the means, along with R² and slope β values reported. “ns” denotes pairwise comparisons between selected lesion-depth bins that were not statistically significant (p > 0.05). Ex vivo depth- and transmurality-sensitive behavior of the selected optical indices. A-D: Index means across 1-mm lesion-depth bins for SDS = 1, 2, 4, and 6 mm. E-H: The same indices across 10% lesion-ratio bins as an indicator for transmurality. Red points represent bin-averaged index values, with variability shown as the mean ± standard deviation (SD) (horizontal and vertical error bars). The blue lines indicate linear fits to the means, along with R² and slope β values reported. “ns” denotes pairwise comparisons between selected lesion-depth bins that were not statistically significant (p > 0.05). Quantitative lesion depth prediction Figure 4 shows a representative example of the quantitative depth-prediction workflow, linking the gross tissue image to the reconstructed lesion-depth ground truth, the raw model output, the lesion-probability mask, the final masked prediction, and the prediction error. The ground-truth map in Fig. 4B shows multiple lesions with distinct depths and spatial distribution. A linear regression model trained on the interpolated multi-SDS optical-index maps generated the raw lesion-depth prediction shown in Fig. 4C, which captured the overall lesion pattern but also produced low-level background predictions outside lesion regions. To suppress these spurious predictions, a logistic regression model was applied to the interpolated multi-SDS lesion-probability maps to generate the lesion-probability mask shown in Fig. 4D. Thresholding this map yielded a binary lesion mask, which was used to set non-lesion predictions to zero, resulting in the final masked depth map shown in Fig. 4E. Compared with the raw regression output, the masked prediction showed improved spatial agreement with the ground truth by preserving lesion-localized depth variation while reducing non-lesion background signal. The corresponding error map in Fig. 4F indicates that the remaining discrepancies were concentrated mainly near lesion boundaries and in regions with the greatest lesion depth. To further evaluate the quantitative depth-prediction model, the predicted and ground-truth lesion depths were compared using binned prediction-versus-actual and Bland-Altman analyses in Supplementary Fig. 3. Feature importance ranking of the multi-SDS model is demonstrated in Supplementary Fig. 4, illustrating the contribution of different SDS in lesion localization and depth estimation. Fig. 4Representative quantitative ex vivo lesion-depth prediction using multi-SDS side-viewing NIRS. A: Gross image of an excised left ventricle with multiple RFA lesions. B: Registered lesion-depth ground truth reconstructed from gross-section measurements. C: Raw lesion-depth prediction generated by the regression model. D: Lesion-probability mask generated by the classification model. E: Final masked lesion-depth prediction obtained after suppressing predictions in non-lesion regions. F: Absolute prediction error relative to the reconstructed lesion-depth ground truth. Representative quantitative ex vivo lesion-depth prediction using multi-SDS side-viewing NIRS. A: Gross image of an excised left ventricle with multiple RFA lesions. B: Registered lesion-depth ground truth reconstructed from gross-section measurements. C: Raw lesion-depth prediction generated by the regression model. D: Lesion-probability mask generated by the classification model. E: Final masked lesion-depth prediction obtained after suppressing predictions in non-lesion regions. F: Absolute prediction error relative to the reconstructed lesion-depth ground truth. Methods Side-viewing NIRS probe design and characterization Multimode fibers (FG200LEA, Thorlabs, Newton, New Jersey, United States) were stripped of their coating and polished to a 45° angle. The angled facet was then coated with a reflective material (301494, Rust-Oleum, Evanston, United States) to promote internal reflection for side-viewing illumination and collection. A probe handle was designed and 3D printed to house the angle-polished fibers. The handle adopted a rigid, paddle-like form factor to approximate the geometry of a commercially available grid-mapping catheter36 to promote surface apposition during mapping. To minimize stray reflections from the housing, the handle was printed using black polylactic acid (PLA) filament. The fibers were then secured to the probe handle using clear epoxy (G14250, Thorlabs). The assembled distal surface was polished to 0.3 μm to provide a smooth tissue-coupling interface. Figure 5A shows a schematic of the side-viewing NIRS probe with an inset highlighting the angled-polished fiber geometry. Figure 5B demonstrated a photograph of the distal end of the device. Fibers a and c were connected to two similar near-infrared (NIR) spectrometers (C9405CB, Hamamatsu, Bridgewater, New Jersey, United States) and served as detector channels for the simultaneous collection of back-scattered light. The spectrometers had a spectral range of 435–1145 nm and a resolution of 0.9 nm. Fibers b and d served as source channels coupled to a broadband light source (HL-2000 HP, Ocean Optics Inc., Dunedin, Florida, United States). When the light source was coupled to fiber b, the spectra collected by fibers a and c corresponded to SDS of 1 and 2 mm, respectively. When coupled to fiber d, the corresponding SDS values were 4 and 6 mm. The spectrometer integration time was set to 30 ms for SDS = 1 and 2 mm to avoid saturation and to 200 ms for SDS = 4 and 6 mm to maintain sufficient signal intensity. The 200 ms exposure time allowed spectra to be collected rapidly and registered to their tracked positions during probe motion, which can reduce the impact of limited slippage on spatial mapping. A separate probe with SDS values of 3, 5, and 7 mm is shown in the Supplementary Figs. 5–10 and Supplementary Table 1. An electromagnetic (EM) tracking probe (TrakStar, Northern Digital Inc., Waterloo, Ontario, Canada) was attached to the back side of the mapping paddle with waterproofed tape to record the spatial location of each acquired spectrum. All acquisition and analysis were performed in MATLAB 2022b and 2023b (The MathWorks Inc., Natick, Massachusetts, United States). Fig. 5Overview of the side-viewing NIRS mapping probe design and EM tracking integration. A: Model of the 3D-printed probe with grooves on the underside to secure optical fibers. Inset: enlarged distal-tip cross-section showing angle-polished fiber (≈ 45°), reflective coating, clear epoxy, and the EM tracking probe mounted on the back side of the paddle. The arrow indicates the representative ray path. B: Photograph of the probe tip demonstrating the positions of the side-viewing multimode fibers and the SDS used for spectral acquisition. Overview of the side-viewing NIRS mapping probe design and EM tracking integration. A: Model of the 3D-printed probe with grooves on the underside to secure optical fibers. Inset: enlarged distal-tip cross-section showing angle-polished fiber (≈ 45°), reflective coating, clear epoxy, and the EM tracking probe mounted on the back side of the paddle. The arrow indicates the representative ray path. B: Photograph of the probe tip demonstrating the positions of the side-viewing multimode fibers and the SDS used for spectral acquisition. Figure 6 depicts the probe characterization through a combination of optical simulation and benchtop beam profiling. A non-sequential Zemax model of the 45°-polished multimode fiber, with 200 μm core size and 0.22 numerical aperture (NA), was constructed to predict the lateral emission and beam divergence, as shown in Fig. 6A. An epoxy block was included because, in the fabricated probe, the angle-polished facet is embedded beneath clear epoxy to simplify assembly, protect the polished surface, and improve mechanical robustness. Beyond this intended role, the epoxy also provided an unexpected optical benefit by acting as a cylindrical lens in front of the facet. It partially diverged the emitted light along the direction that is perpendicular to the fiber, producing a more circular spot profile. Beam profiles were then measured experimentally at working distances of 0 to 5 mm from the probe tip, as displayed in Fig. 6B-E. Gaussian fits to the measured intensity profiles (red curves) were used to extract the full-width half-maximum (FWHM) values along the x- and y-axes. The FWHM increases from ~ 1.0 mm at 0 mm to ~ 3.0 mm at 5 mm, indicating controlled beam expansion with distance and good agreement with the simulated emission pattern. Based on the measured divergence, the output NA of the side-viewing fiber was estimated as 0.20 along the x-axis and 0.18 along the y-axis. Fig. 6Characterization of the side-viewing fiber through simulation and experiment. A: Zemax simulation of the side-viewing fiber, showing the diverging light pattern at the angled tip. An epoxy block (n = 1.66) was added to replicate the real probe. B - E: Experimental beam profiles of a side-viewing fiber at varying distances (0.0 to 0.5 mm) from the probe tip, demonstrating good agreement with the simulation. Blue lines represent the measured intensity profiles, while red lines indicate Gaussian fits to the experimental data. Characterization of the side-viewing fiber through simulation and experiment. A: Zemax simulation of the side-viewing fiber, showing the diverging light pattern at the angled tip. An epoxy block (n = 1.66) was added to replicate the real probe. B - E: Experimental beam profiles of a side-viewing fiber at varying distances (0.0 to 0.5 mm) from the probe tip, demonstrating good agreement with the simulation. Blue lines represent the measured intensity profiles, while red lines indicate Gaussian fits to the experimental data. Depth sensitivity of multiple SDS Figure 7 quantifies how increasing SDS shifts the sensitivity profile toward deeper tissue using Monte Carlo photon-transport simulations implemented in the pyxopto Python package39. Illumination and collection were modeled with multimode fibers using an effective NA of 0.19, selected as the average of the experimentally measured NA values reported above. Tissue optical properties were set to an absorption coefficient of µa = 0.5 cm− 1 and a reduced scattering coefficient of µs’ = 20 cm− 1, representing typical values for myocardial tissue40,41. In Fig. 7A, voxel-wise maps of detected-photon pathways in the XZ cross-sectional plane are shown on a log scale for SDS ranging from 1 to 6 mm. The maps indicate the relative likelihood that photons ultimately collected by the detector fiber passed through each voxel. The red dashed line denotes the central axis used to extract the depth profiles, and the horizontal dashed lines indicate penetration depth at 3 dB (yellow), 10 dB (green), and 15 dB (magenta). The blue curves on the right of each panel show the central-axis distributions, with corresponding log-normal fits overlaid in red29. Figure 7B summarizes the penetration depth extracted from the dashed-line thresholds in Fig. 7A, with the colored rectangles and circles corresponding to each attenuation level. The results show progressively deeper sampling with increasing SDS. Fig. 7Monte-Carlo simulation of SDS-dependent penetration depth. A: Detected-photon pathway maps for SDS from 1 to 6 mm, shown on a log scale. The colorbar represents the voxel-wise contribution to the detected signal. Triangles mark the positions of the source and detector fibers. The red dashed line marks the central axis; horizontal dashed lines indicate 3 dB (yellow), 10 dB (green), and 15 dB (magenta) penetration depths. Blue curves show the central-axis profiles and red curves the corresponding log-normal fits. B: Penetration depth versus SDS, showing deeper sampling with increasing SDS. Monte-Carlo simulation of SDS-dependent penetration depth. A: Detected-photon pathway maps for SDS from 1 to 6 mm, shown on a log scale. The colorbar represents the voxel-wise contribution to the detected signal. Triangles mark the positions of the source and detector fibers. The red dashed line marks the central axis; horizontal dashed lines indicate 3 dB (yellow), 10 dB (green), and 15 dB (magenta) penetration depths. Blue curves show the central-axis profiles and red curves the corresponding log-normal fits. B: Penetration depth versus SDS, showing deeper sampling with increasing SDS. Experimental workflow and labeling Experiments were conducted on 11 healthy fresh porcine hearts (Green Village Packing Co., Green Village, New Jersey, United States) following the workflow summarized in Fig. 8. The left ventricle (LV) was excised from the atria and right ventricle, opened with an incision along the septum, and flattened to expose the epicardial surface for mapping. The specimen was positioned on an electrosurgical grounding pad and maintained in a 37 °C phosphate-buffered saline (PBS) bath. Irrigated RFA lesions were then created on the epicardium using 30 to 40 W for 20 to 100 s, with continuous irrigation at 17 mL/min delivered to the catheter tip by a clinical pump (Biosense Webster, Irvine, California, United States). Ablation settings were deliberately varied to generate lesions spanning a range of depths across heterogeneous myocardial thicknesses, while minimizing the risk of excess heating effects such as steam pots, resulting in 4 to 20 lesions per ventricle (a total of 133 lesions). After lesion creation, each ventricle was transferred to a separate mapping chamber and pinned to a corkboard. Gross photographs were acquired to enable the subsequent registration of spectral sampling locations to the tissue image32,33. PBS was periodically sprayed during mapping to maintain a moist epicardial surface. NIRS spectra were collected using two side-viewing probes with SDS values spanning from 1 to 7 mm. Then, the ventricles were stained with 1% 2,3,5-triphenyl-2 H-tetrazolium chloride (TTC) at 0–4 °C for over 24 h to enhance lesion visibility for lesion depth assessment. Adipose regions were identified on the gross epicardial surface based on their visible appearance in the tissue photographs during mapping and registration. All ablation and NIRS mapping procedures were completed within two days postmortem. After TTC staining, each lesion was bisected through its center and imaged on a clean, flat background. Photographs were acquired using a smartphone camera, with a millimeter ruler included in each frame to provide a physical scale. The lesion center was manually identified, and lesion depth was measured as a function of radial distance by converting pixel distances to physical values along directions perpendicular and parallel to the epicardial surface. Total tissue thickness per lesion was also measured based on the physical scale for transmurality analysis. The distance measurements were rounded to the nearest 0.1 mm increment for consistency with the practical precision of the photographic imaging setup. Assuming approximate central symmetry, the lesion depth measurements were used to reconstruct a 3D depth distribution for each ablation site. Lesion centers were also manually annotated on the gross epicardial image. Using the inverse of the physical-to-image registration, each image pixel was mapped to its corresponding physical location, allowing the per-lesion 3D depth reconstructions to be assembled into a lesion-depth ground-truth map. To account for the reduced spatial resolution associated with increasing SDS, the ground-truth map was smoothed with a Gaussian filter. The kernel width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{px}$$\end{document} of the filter was set by converting the SDS to pixel units using the approximate pixel-to-millimeter conversion factor, computed as the Euclidean norm of the first two elements of the first column of the projective transformation matrix returned by MATLAB’s fitgeotrans command. The standard deviation was then chosen such that ± 3σ spanned the kernel. Finally, the Gaussian standard deviation was set as1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}\sigma\:={\lceil\frac{{W}_{px}-1}{6}\rceil}_{odd}\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{px}$$\end{document} denotes the kernel width in pixel units, and the “odd” subscript indicates that σ was increased to the next odd integer to ensure an odd-sized kernel for symmetric smoothing about the center pixel. Fig. 8Workflow for LV lesion creation, NIRS mapping, and ground-truth generation. The left ventricle was dissected, and irrigated RFA lesions were created in 37 °C PBS, after which ex vivo NIRS mapping was performed using the EM-tracked side-viewing probe. Lesions were subsequently TTC-stained, bisected, and photographed next to a ruler. From these images, lesion depth was quantified as a function of radial distance from the lesion center, assuming the lesions were centrally symmetric. Finally, the measurements were assembled into ground-truth lesion-depth maps. Workflow for LV lesion creation, NIRS mapping, and ground-truth generation. The left ventricle was dissected, and irrigated RFA lesions were created in 37 °C PBS, after which ex vivo NIRS mapping was performed using the EM-tracked side-viewing probe. Lesions were subsequently TTC-stained, bisected, and photographed next to a ruler. From these images, lesion depth was quantified as a function of radial distance from the lesion center, assuming the lesions were centrally symmetric. Finally, the measurements were assembled into ground-truth lesion-depth maps. NIRS signal analysis To generate normalized relative reflectance spectra, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{rel}\left(\lambda\:\right)$$\end{document}, a dark spectrum was first subtracted from the raw measurements to remove background and detector noise, after which the spectra were corrected for wavelength-dependent system response by referencing a 99% diffuse reflectance standard (Spectralon, LabSphere, North Sutton, New Hampshire, United States). Spectra were then scaled using a polydimethylsiloxane (PDMS) phantom at a wavelength with known optical properties, band-limited to the 500–1100 nm range, low-pass filtered with a 0.05π rad/sample cutoff frequency, and normalized by the AUC over the truncated range, reducing amplitude variability. Derivative spectra \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{rel}^{{\prime\:}}\left(\lambda\:\right)$$\end{document} were computed from the resulting relative spectra. Measurements that saturated the spectrometers were excluded, and inadequate contact points were rejected by removing spectra whose maximum intensity fell below 40% of the peak in the PDMS phantom reference spectrum. After quality filtering, spectra from 8 hearts with 105 lesions were retained for 1 and 2 mm SDS values, and 9 hearts with 109 lesions for 4 and 6 mm SDS values. For each of the smaller SDS values (1 and 2 mm), the analyzed dataset comprised 2922 spectra from lesions > 4 mm, 2198 spectra from lesions ≤ 4 mm, 1901 adipose spectra, and 6504 spectra from untreated epicardial muscle. For each of the larger SDS values (4 and 6 mm), the corresponding counts were 3099, 3379, 2161, and 6942, respectively. Optical index development and statistical analysis Optical indices were developed separately for each SDS using an exhaustive grid-search method33,42 to identify compact spectral metrics that best discriminate key tissue types, including lesion versus non-lesion (termed “lesion optical index”, or LOI), lesions ≤ 4 mm versus lesions > 4 mm (termed “depth contrast index”, or DCI), and adipose versus non-adipose tissue (termed “adipose contrast index”, or ACI). Further details about the exhaustive grid-search method can be found in the Supplementary Information. For DCI development, lesions were divided into two groups (lesions of depth > 4 mm and lesions ≤ 4 mm) for practical binary comparison and visualization, while the main depth analyses used continuous depth values and 1 mm depth bins. The grid-search approach was also adapted to identify indices that track lesion depth rather than class labels, termed the lesion depth index (LDI). Instead of selecting candidates by maximizing the area under the ROC curve for binary classification, candidate indices computed from lesion spectra were regressed against the measured lesion depth using a linear model, and indices were ranked by the R2. This depth-sensitivity search was performed using two depth definitions: continuous lesion depth reported at 0.1 mm resolution and lesion depth rounded into 1 mm bins. When the same index was optimal under both definitions, a single LDI was retained; otherwise, separate indices were selected for each depth definition. The complete set of optical indices identified for each separation is provided in the Supplementary Information, where superscripts denote SDS and subscripts are used only for LDIs to enumerate multiple depth-sensitive indices; one LOI, one ACI, and one DCI were selected per SDS, whereas up to six LDIs were retained and indexed by the subscript. Statistical analyses were performed to evaluate the discriminative and depth-sensitive behavior of the optical indices across SDS values. For each source-detector separation, the classification performance of individual LOIs and ACIs was assessed using ROC curves and the corresponding AUC for lesion versus non-lesion and adipose versus non-adipose discrimination. To assess lesion depth sensitivity, LOIs, DCIs, and LDIs for each SDS were grouped into 1-mm lesion depth bins and analyzed using one-way analysis of variance (ANOVA) with multiple comparison testing to evaluate differences among the depth bins, with a significance threshold of p < 0.05. Additionally, mean index values were fitted against binned lesion depths using a linear model, and the slope and R2 were compared along with the number of lesion depth bin pairs that were not statistically significant. The optimal indices selected for each SDS are listed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{LDI}_{2}^{1\:mm}=\sum\limits_{605\:nm}^{655\:nm}{R}_{rel}\left(\lambda\:\right)\end{array}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{LDI}_{2}^{1\:mm}=\sum\limits_{915\:nm}^{965\:nm}{R}_{rel}\left(\lambda\:\right)\end{array}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{LDI}_{4}^{4\:mm}=\frac{{R}_{rel}\left(830\:nm\right)}{{R}_{rel}\left(920\:nm\right)}\end{array}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}{DCI}^{6mm}=\frac{{R}_{rel}\left(925\:nm\right)}{{R}_{rel}\left(825\:nm\right)}\end{array}$$\end{document} The selected optimal indices were further assessed for sensitivity to lesion transmurality. Lesions were grouped into 10% lesion ratio bins, and one-way ANOVA with multiple comparison testing was performed across bins. Signal processing was carried out in MATLAB 2022b (The MathWorks Inc., Natick, Massachusetts, United States), and statistical testing was performed in MATLAB 2022b and GraphPad Prism 10.3.1 (GraphPad Software, United States). Lesion classification and depth mapping RF classifiers were developed to distinguish lesion from non-lesion tissue using features derived from the NIRS measurements. Data were analyzed independently, and RF models were trained separately for each SDS. For each SDS, the input feature set consisted of the optical indices described in the sections above. Model evaluation used a leave-one-heart-out cross-validation scheme to test generalization across specimens, where all data from one heart were held out for testing, while the remaining hearts were used for training in each fold. This strategy prevents data leakage across hearts and reduces variability associated with random train-test splits in conventional k-fold cross-validation. Quantitative lesion-depth prediction was performed by combining SDS-specific interpolated optical-index maps with a lesion-probability mask. For each map and each SDS (1, 2, 4, and 6 mm), the selected depth-sensitive optical indices (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{LDI}_{2}^{1\:mm}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{LDI}_{2}^{2\:mm}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{LDI}_{4}^{4\:mm}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{DCI}^{6\:mm}$$\end{document}) and the RF lesion output (as shown in Supplementary Fig. 2) were interpolated separately onto the registered epicardial image grid using natural-neighbor interpolation. The interpolated maps were then Gaussian-smoothed using the SDS-matched kernel width, defined in Eq. 1, set in pixel units from the image-to-physical registration. A pixel-wise dataset was assembled across 8 maps by stacking the four interpolated optical-index maps (one for each SDS) as regression features and the four interpolated RF-output maps (one for each SDS) as classification features. Using 6 maps for training and 2 for testing, a linear regression model was trained on lesion pixels only to predict continuous lesion depth, while a logistic regression model was trained on all valid pixels to estimate lesion probability. For each training and test pixel, the linear model generated a raw depth prediction, and the logistic model generated a lesion-probability map. The probability map was thresholded at 0.5 to form a binary lesion mask, and predicted depths in pixels classified as non-lesion were set to zero to obtain the final masked depth map. Representative examples of the gross image, registered ground truth, raw regression prediction, lesion-probability mask, masked prediction, and absolute prediction error are shown in Fig. 4. All model development and map visualization were carried out in MATLAB 2022b. Discussion In this study, we demonstrated a fiber-based, side-viewing NIRS mapping approach for depth-sensitive characterization of epicardial RFA lesions on the porcine left ventricle. By integrating a compact side-viewing geometry with multiple SDS, the system integrates lesion localization to depth-related optical contrast and produces spatial maps that agree with independently reconstructed lesion-depth ground truth. Optical characterization showed that embedding 45°-polished fiber tips beneath clear epoxy ensures robustness while shaping the output beam into a stable circular profile suitable for close-range mapping on curved epicardial surfaces. Monte-Carlo modeling supports the multi-SDS strategy by showing that increasing SDS shifts the photon-sampling distribution deeper. Separation-dependent spectral differences across lesion, untreated epicardial muscle, and adipose tissue motivated an optical-index framework for both tissue identification and lesion depth assessment. Across SDS values, single indices provided strong discrimination between lesion versus non-lesion and adipose versus non-adipose, and depth-sensitive indices varied monotonically with both lesion depth and lesion ratio, enabling depth-related contrast after lesion localization. RF classification produced a practical lesion mask for downstream mapping and performed better at smaller separations. The consistently higher specificity than sensitivity is consistent with class imbalance favoring non-lesion rejection, and the weaker performance at larger SDS is consistent with broader sampling that can reduce local contrast at lesion boundaries. To further evaluate the practical value of integrating information across separations, we also performed a quantitative depth-prediction analysis using interpolated multi-SDS optical-index maps together with a lesion-probability mask. The representative results showed that this framework could recover both lesion location and relative depth variation, while masking reduced non-lesion background signal and improved agreement with the reconstructed ground truth. These results are consistent with the complementary roles of the different separations, with smaller SDS emphasizing local lesion contrast and larger SDS providing additional subsurface information for depth estimation. These findings further support the value of the multi-SDS approach not only for depth-sensitive qualitative mapping but also for quantitative lesion-depth prediction. Compared with earlier NIRS catheter work focused on tracking lesion progression and estimating depth in forward-viewing or ablation-integrated configurations30,31, the approach in this paper emphasizes post-ablation spatial mapping with a side-viewing, multi-SDS geometry to visualize heterogeneity across multiple lesions on the epicardial surface. Recent work has also demonstrated robust NIRS-based lesion mapping in atrial settings32, supporting the broader feasibility of catheter-based optical lesion assessment. More generally, other optical modalities, like OCT30,43, have shown lesion visualization and boundary detection capabilities, with tradeoffs in penetration depth, hardware complexity, and field of view. Within this landscape, depth-sensitive NIRS mapping offers a balance of simplicity and depth reach. Clinically, epicardial ablation outcomes are sensitive to lesion depth and completeness, but verifying the adequate formation of lesions during procedure remains challenging. A depth-sensitive mapping layer that can be overlaid on a lesion-localization mask could provide practical feedback on the depth of the ablation lesion and help operators identify areas that may require additional energy delivery or alternative strategies. The potential to package a side-viewing, fiber-based architecture into a slim, elongated form factor also supports a realistic path toward integration with contemporary electrophysiological mapping catheters. Moreover, a maximum lesion ratio of 0.6 is clinically meaningful because conventional unipolar RF lesions are typically on the order of 6–8 mm deep44, which corresponds to a lesion ratio of approximately 0.6 in ventricular walls of about 10–12 mm thickness45,46. In thicker or hypertrophied ventricles47, the same ratio would still represent a subtransmural lesion. In addition, the multi-SDS configuration in this study was developed to assess tissue and therapeutic changes over the approximate depth range (5–8 mm) supported by Monte Carlo analysis, which is relevant for epicardial lesion assessment. Accordingly, the side-viewing multi-SDS geometry may also be useful in other organ systems when the target therapy or tissue change lies within a similar depth range and probe contact is feasible. However, in tissues with different absorption, scattering, thickness, and surface topography, the optimal SDS range, probe-tissue interface, and spectral indices would still need to be re-optimized and validated for the specific application. Acute in vivo edema and perfusion changes may alter scattering and absorption relative to static ex vivo tissue48–50. However, prior cardiac NIRS studies from our group suggest that these effects do not eliminate lesion contrast in vivo. In NIRS-integrated RF catheter work31, a model developed from ex vivo lesion data showed robust translational consistency in vivo pig studies. In separate atrial mapping experiments32, lesion contrast also remained reliable in blood and at oblique angles, and lesion maps in PBS and blood showed strong agreement. These findings suggest that motion, blood, edema, and perfusion changes may introduce variability and reduce quantitative accuracy, but lesion-versus-non-lesion classification can still remain feasible. This study has several limitations. The experiments were performed ex vivo on porcine ventricles, so optical contrast and classifier behavior may differ in vivo due to blood, motion, and temperature-dependent tissue changes. Progressive heart failure remodeling could alter the spectral readout through fibrosis, extracellular matrix remodeling, microvascular rarefaction, metabolic shifts, and potentially increased epicardial adipose, all of which can change scattering and absorption independent of ablation51,52. As a result, a depth model trained on relatively static ex vivo tissue may misattribute substrate-related spectral changes to lesion depth and show reduced specificity in remodeled hearts. At the same time, our prior NIRS studies30,33,53 indicate that adipose and other epicardial substrates can be detected spectroscopically, suggesting that future translation to diseased myocardium will likely benefit from substrate-aware or multiclass modeling rather than a lesion-only framework. Additionally, the present study was limited to RFA, as it remains unclear whether the NIRS contrast mechanisms and indices identified here would transfer to pulsed field ablation, and this would require dedicated investigation. Future work will focus on transitioning from index-based trends to quantitative depth prediction by developing regression models that leverage a combined, multi-SDS dataset, allowing complementary penetration depths to be fused within a single machine-learning framework for improved accuracy. These efforts will be supported by expanding the dataset to a larger sample size to improve statistical power, better capture inter-subject variability, and enable more robust model training and validation. Stable optical coupling on a curved, wet, beating epicardial surface remains an important challenge for in vivo translation. Although the paddle-like geometry used here was selected to promote surface apposition, future designs will require more flexible spline-like structures36 for mapping along curved surfaces. In addition, future translation will also rely on contact assessment, as prior NIRS-integrated catheter work31 has shown that a contact optical index can distinguish contact from non-contact, both ex vivo and in vivo. Conclusion We introduced a fiber-based, side-viewing, multi-SDS NIRS mapping probe and demonstrated its ability to provide depth-sensitive assessment of epicardial RFA lesions on the porcine left ventricle. Across separations, compact optical indices enabled the identification of lesions and adipose tissue and showed monotonic trends with lesion depth and transmurality. The multi-SDS quantitative analysis provided lesion-depth predictions that recovered lesion location and relative depth variation with improved agreement with the reconstructed ground truth after lesion-probability masking. Overall, these results support side-viewing multi-SDS optical mapping as a promising auxiliary intraoperative tool for lesion validation and depth assessment. Future work will focus on in vivo validation and further refinement of the quantitative depth-prediction framework. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1