Femtosecond label-free imaging: A rapid and reliable alternative for intraoperative pathological assessment in thoracic oncology Authors: Hao Yin, Fangyi Liu, Wendi Zhu, Wenlong Yu, Kunbo Zhang, Rongkui Luo, Fenghao Sun, Yunlong Wang, Boxue Zhang, Bingwei Xu, Xin Zhu, Mingxiang Feng, Lijie Tan, Ming Li Published: 2026 ABSTRACT AbstractObjectiveThe study objective was to evaluate femtosecond label-free imaging combined with artificial intelligence as a rapid alternative to frozen-section pathology for intraoperative assessment in thoracic oncologic surgery.MethodsThis prospective study enrolled 144 patients undergoing resection for thoracic tumors at Shanghai Zhongshan Hospital (June to December 2025). Fresh specimens (259 lung, 96 esophageal) underwent femtosecond label-free imaging scanning using ultrashort laser pulses to generate multiple nonlinear optical signals (third harmonic generation, second harmonic generation, 2-photon and 3-photon fluorescence) without sectioning or staining. Deep learning algorithms were developed for lung tumor classification and exploratory high-risk feature prediction in esophageal squamous cell carcinoma. To avoid data leakage, specimen-level splits were adopted for lung tumor classification and patient-level splits for esophageal squamous cell carcinoma high-risk feature prediction. The models were constructed based on the pretrained UNI v1 foundation model combined with an attention-based multiple-instance learning framework, with data augmentation (flipping and rotation) applied during training. Performance was evaluated against conventional histopathology using receiver operating characteristic curve analysis, with data partitioned into training, validation, and test sets at a 7:1:2 ratio.ResultsFemtosecond label-free imaging demonstrated significant time advantage over frozen-section analysis (median 5.4 vs 36.3 minutes, P < .001). Morphological concordance with hematoxylin–eosin histology was achieved across tissue types. Artificial intelligence–based tumor classification achieved a mean area under the curve of 0.953 for lung specimens. Depth scanning at 3-μm intervals (up to 90 μm) enabled volumetric margin assessment, revealing depth-dependent tumor distribution variations not captured by single-plane frozen sections. Exploratory models predicted high-risk features in esophageal squamous cell carcinoma with a mean area under the curve of 0.653 to 0.766 for lymph node metastasis, perineural invasion, and lymphovascular invasion.ConclusionsFemtosecond label-free imaging provides rapid, morphologically concordant pathological assessment with significant workflow advantages over conventional frozen-section analysis. Depth-scanning capability addresses sampling limitations inherent to single-plane evaluation. Exploratory artificial intelligence models demonstrate feasibility for intraoperative prediction of high-risk pathological features in esophageal cancer; however, these results are preliminary and require validation in larger, multicenter cohorts before clinical deployment.Key Words: esophageal cancer, femtosecond label-free imaging, frozen section, intraoperative diagnosis, lung cancer, multimodal optical imaging FLI imaging shows morphological concordance with HE in lung and esophageal tumors. Central Message FLI with AI provides rapid, accurate intraoperative pathology in thoracic surgery. Perspective FLI with AI achieves accurate tumor diagnosis in less than 6 minutes. Depth scanning enables comprehensive volumetric margin assessment critical for sublobar resections and reveals tumor distributions missed by single-plane evaluation. This platform offers standardized real-time surgical decision support with early risk stratification in thoracic oncology. Intraoperative frozen-section pathology remains the standard for real-time histological assessment during thoracic oncologic surgery, yet presents significant limitations affecting surgical decision-making. The process consumes valuable operative time and depends heavily on technical expertise and workflow coordination, and diagnostic performance may vary across institutions due to differences in specimen handling and clinical experience.1,2 Under time constraints, pathologists often examine only 1 or 2 sections per specimen, potentially missing critical findings in heterogeneous tumors; in thoracic oncology, sampling-related underestimation has been reported as a key contributor to discordance between frozen and permanent sections.2,3 Quality standardization across centers remains challenging due to freezing artifacts, tissue fragmentation, and compression that can compromise diagnostic confidence and occasionally necessitate deferral to permanent sections. These limitations directly impact clinical decisions in thoracic surgery, where accurate margin assessment determines resection extent, balancing oncological adequacy against organ function preservation.1,3 Femtosecond label-free imaging (FLI) represents a fundamentally different approach. Using ultrashort laser pulses, FLI generates multiple nonlinear optical signals from endogenous biomolecules without requiring tissue sectioning, fixation, or staining. Related nonlinear optical microscopy frameworks have been shown to yield histomorphology that is interpretable in a pathology-like manner.4,5 The technology simultaneously captures third harmonic generation (THG), 2-photon fluorescence emission (2PFE), second harmonic generation (SHG), and 3-photon fluorescence emission (3PFE) to create high-resolution images revealing tissue architecture and cellular morphology, consistent with prior evidence that third THG/SHG and multiphoton signals can highlight cellularity and stromal organization without conventional staining.6,7 Unlike frozen sections sampling a single tissue plane, FLI scans through tissue depth, generating volumetric datasets that more comprehensively represent the specimen; depth-resolved “slide-free” imaging has been proposed as a practical route to reduce plane-selection bias in rapid pathology workflows. Related applications of FLI and nonlinear label-free optical imaging have been reported in several tumor types,8, 9, 10 supporting their potential value in oncologic tissue assessment. However, the possible clinical application and significance of this approach in thoracic tumors remain to be further explored. Combining FLI with artificial intelligence (AI) offers additional advantages. AI-based interpretation can reduce interobserver variability and facilitate standardized diagnosis across centers, a goal that has been emphasized in computational pathology literature and validated in large-scale, whole-slide, deep-learning studies.11 Once validated, AI algorithms can be deployed universally, potentially supporting consistent care delivery and multicenter evaluation with harmonized decision thresholds. We hypothesized that FLI combined with AI could (1) provide rapid, accurate pathological assessment comparable to frozen-section analysis with advantages in workflow efficiency, tissue preservation, depth sampling, and standardization; and (2) determine whether FLI images encode microstructural correlates of high-risk pathological features—including lymphovascular invasion, perineural invasion, and lymph node metastasis—as an exploratory step toward characterizing the diagnostic capability boundaries of the platform. Analogous to the development of frozen-section pathology, which required initial demonstrations of diagnostic feasibility before clinical utility could be established, determining whether FLI can capture these features at all is a necessary scientific precursor to future investigations of clinical impact. We acknowledge that high-risk features with more direct intraoperative decision relevance in non–small cell lung cancer—including spread through air spaces and visceral pleural invasion, whose intraoperative identification may support conversion from sublobar to lobar resection—represent a more immediately actionable target; parallel predictive models for these end points are under active development and will be reported in a subsequent study. Materials and Methods Study Design and Sample Collection This prospective study enrolled consecutive patients undergoing surgical resection for suspected thoracic malignancies at Shanghai Zhongshan Hospital between June 1, 2025, and December 31, 2025. After resection, fresh tumor specimens were promptly delivered for FLI scanning in parallel with intraoperative frozen-section pathological assessment. On completion of FLI and frozen-section evaluation, the tissue was subsequently processed for routine permanent histopathological evaluation. The study protocol was approved by the Zhongshan Hospital Institutional Review Board (B2025-222R, approved May 16, 2025), and all participants provided written informed consent. For each patient, specimens were collected from 1 or more of the following categories: (1) primary tumor tissue, (2) corresponding normal/uninvolved tissue from the same resection specimen, (3) surgical resection margin tissue, and (4) regional lymph nodes as applicable. Femtosecond Label-Free Imaging System Images were acquired using an FLI microscopy system (FI-100, Femtosecond Research Center [Guangzhou] Co, Ltd). The system used sub-20 femtosecond laser pulses with a broad spectral range (900-1200 nm) at a repetition rate of 12.5 MHz. Pulses were focused through a high numerical aperture water immersion objective (UAPON 40XW340, Olympus; NA 1.15) to interact with endogenous biomolecules in fresh tissues without labeling or sectioning. All specimens in this study were scanned fresh, immediately after resection. Although the system is technically capable of imaging fixed tissues, fixation-related changes in protein cross-linking and autofluorescence may alter signal characteristics, and fixed-tissue imaging was not performed in this study. Multiple nonlinear optical modalities were generated simultaneously: THG, 2PFE, SHG, and 3PFE. Signals were collected through a multi-channel detection system. Images from individual channels were merged to generate composite pseudo-color images. The FLI system provides a single-frame field of view of 360 × 360 μm2. For diagnostic assessment of lung cancer lesions, a region of interest of 2500 × 2500 μm2 was scanned by acquiring and mosaicking consecutive frames. Sequential depth images were acquired at 3-μm intervals through tissue thickness (up to 90 μm), generating volumetric datasets. Three-dimensional reconstructions were created from these depth-series images. For margin assessment, FLI was performed on fresh resection margin specimens obtained at the time of surgery. FLI-derived margin evaluations were subsequently compared with final permanent-section histopathological results, which served as the reference standard for true margin status. Specimen Preparation and Femtosecond Label-Free Imaging Scanning Workflow The entire operational process is demonstrated in Figure 1, A and Video 1: (1) Specimen preparation: Fresh tissue samples were rinsed with normal saline and trimmed to approximately 0.5 cm × 0.5 cm × 0.5 cm. FLI specimens were collected in parallel with, but independent from, frozen-section specimens. Each sample was placed in a confocal culture dish with care taken to eliminate air bubbles between the sample and dish bottom. The sample was secured with a holder, a positioning image was captured, and a region of interest was selected for scanning. (2) FLI scanning: The sample holder was loaded into the FLI system, and after being initiated, the system performed automatic focusing followed by image acquisition. The number of specimens scanned per patient varied according to the intraoperative sampling plan (primary tumor, resection margin, or regional lymph nodes). (3) FLI-AI analysis: Scanning data were imported into the FLI-AI analysis system, which automatically identified tissue characteristics, classified benign versus malignant tissue, and output a diagnostic result. Morphological concordance between FLI (THG: yellow, SHG: green, 3PFE: magenta) and conventional histopathology. A, Schematic workflow of fresh tissue acquisition, FLI scanning, AI-based analysis, and parallel conventional histopathological processing. B, Lung adenocarcinoma: FLI shows malignant glandular architecture and nuclear atypia (triangles, tumor cells with increased nuclear-cytoplasmic ratio and indistinct nuclear boundaries), with corresponding HE-stained section. C, Adjacent normal lung tissue: FLI shows preserved alveolar structures (white dashed box), with corresponding HE-stained section. D, ESCC: FLI shows cellular pleomorphism and architectural disorganization (triangles, tumor cells; pentagrams, keratin pearls), with corresponding HE-stained section. E, Adjacent normal esophageal mucosa: FLI shows intact stratified squamous epithelium, with corresponding HE-stained section. FLI, Femtosecond label-free imaging; AI, artificial intelligence; LUAD, lung adenocarcinoma; H&E, hematoxylin–eosin; ESCC, esophageal squamous cell carcinoma. Artificial Intelligence Algorithm Development Deep learning algorithms were developed for lung tumor classification and exploratory prediction of high-risk pathological features in esophageal squamous cell carcinoma (ESCC). For lung tumor classification, FLI images were derived from 259 lung specimens and labeled according to the corresponding final histopathological diagnosis, with pathologist review of paired hematoxylin–eosin (HE) sections used to confirm tumor and nontumor regions. For ESCC high-risk feature prediction, FLI images were derived from 96 esophageal specimens from 43 patients, and ground-truth labels for lymph node metastasis (LNM), perineural invasion (PNI), and lymphovascular invasion (LVI) were obtained from the final pathology reports. For ESCC high-risk feature prediction, data partitioning was performed at the patient level because LNM, PNI, and LVI were patient-level pathological labels; all specimens and image patches from the same patient were kept within the same fold. This strategy was used to prevent data leakage and to avoid overestimation of model performance. For each FLI specimen, background regions were removed using foreground segmentation, and the remaining foreground areas were divided into patches. Each patch was augmented during training by flipping and rotation, encoded using the pretrained UNI v1 foundation model, and grouped with other patches from the same specimen as a multiple-instance learning bag. The ABMIL classifier uses a gated attention mechanism to weight informative patches before generating a specimen-level prediction through a linear classifier. The lung tumor model was trained with cross-entropy loss and a bag-level batch size of 1. Data were split at the specimen level into training, validation, and test sets at a ratio of 7:1:2. For ESCC, the same FLI-derived features were used in exploratory models to predict lymph node metastasis, perineural invasion, and lymphovascular invasion with patient-level data partitioning. Attention mapping using gradient-weighted class activation mapping was applied to visualize image regions contributing to model predictions. Statistical Analysis Morphological concordance between FLI and HE histology was assessed by independent side-by-side review of paired images conducted by 2 evaluators (a thoracic pathologist and a thoracic surgeon experienced in FLI interpretation). Concordance was defined as agreement on tissue classification (malignant vs benign) and on the presence of key diagnostic morphological features as described for each tissue type. Discrepancies between reviewers were resolved by consensus review. Continuous variables are presented as median with interquartile range (IQR). Imaging time comparisons between FLI and frozen-section analysis were performed using the Mann–Whitney U test. Diagnostic performance metrics were calculated with 95% CIs. Receiver operating characteristic (ROC) curves were generated for AI prediction models. Results Feasibility and Workflow Integration A total of 355 specimens from 144 patients were included in this study, comprising lung and esophageal surgical samples. For lung specimens, 259 tissues were obtained from 101 patients, including 62 invasive adenocarcinomas (61.4%), 13 minimally invasive adenocarcinomas (12.9%), 13 lung squamous cell carcinomas (12.9%), 4 pulmonary metastases (PMs, 4.0%), 7 benign tumors (6.9%), and 2 other lesions (2.0%). The majority of lung cancer cases were early-stage disease, with pT1b representing the most common pathological stage (50.0%). Nodal metastasis was present in 12.5% of cases, with pN1 in 8.0% and pN2 in 4.5%. High-risk pathological features included LVI in 8.2% of cases, PNI in 1.4%, spread through air spaces (STAS) in 23.3%, and PNI in 11.5%. Detailed clinicopathological characteristics are summarized in Table 1. Table 1Clinicopathological characteristics for lung casesFeaturesLung casesn (%)101 (100%)Histology (%) BT7 (6.9) IAC62 (61.4) LSCC13 (12.9) PM4 (4.0) MIA13 (12.9) Other2 (2.0)pT (%) 1a8 (9.1) 1b44 (50.0) 1c8 (9.1) 1mi13 (14.8) 214 (15.9) 31 (1.1)pN (%) 077 (87.5) 17 (8.0) 24 (4.5)LVI (%) 067 (91.8) 16 (8.2)PNI (%) 069 (98.6) 11 (1.4)STAS (%) 056 (76.7) 117 (23.3)Pleural invasion (%) PL069 (88.5) PL19 (11.5)Clinical and pathological features of lung specimens examined in this study. Histological subtypes include BT, IAC, LSCC, PM, MIA, and other subtypes. Pathological staging (pT and pN), LVI, PNI, STAS, and PL status are presented as percentages of evaluable cases. BT, Benign tumor; IAC, invasive adenocarcinoma; LSCC, lung squamous cell carcinoma; PM, pulmonary metastasis; MIA, minimally invasive adenocarcinoma; LVI, lymphovascular invasion; PNI, perineural invasion; STAS, spread through air spaces; PL, pleural invasion. Clinicopathological characteristics for lung cases Clinical and pathological features of lung specimens examined in this study. Histological subtypes include BT, IAC, LSCC, PM, MIA, and other subtypes. Pathological staging (pT and pN), LVI, PNI, STAS, and PL status are presented as percentages of evaluable cases. BT, Benign tumor; IAC, invasive adenocarcinoma; LSCC, lung squamous cell carcinoma; PM, pulmonary metastasis; MIA, minimally invasive adenocarcinoma; LVI, lymphovascular invasion; PNI, perineural invasion; STAS, spread through air spaces; PL, pleural invasion. For esophageal tumors, 96 specimens were collected from 43 patients, all with squamous cell carcinoma (ESCC, 100%). R0 resection was achieved in all cases. Most cases were staged as pT3 (46.5%), followed by pT2 (30.2%), pT1b (18.6%), pT0 (2.3%), and pT4a (2.3%). Nodal metastasis was observed in 44.2% of patients, with pN1 in 20.9%, pN2 in 18.6%, and pN3 in 4.7%. LVI was present in 46.5% of cases, and PNI was present in 48.8% of cases. Seven patients (16.3%) received neoadjuvant therapy before surgery, with tumor regression grades ranging from tumor regression grade 0 (14.3%) to 3 (42.9%). Corresponding pathological features are provided in Table 2. Table 2Clinicopathological characteristics for esophageal casesFeaturesEsophageal casesn (%)43 (100.0)ESCC (%)43 (100.0)R0 resection (%)43 (100.0)pT (%) 01 (2.3) 1b8 (18.6) 213 (30.2) 320 (46.5) 4a1 (2.3)pN (%) 024 (55.8) 19 (20.9) 28 (18.6) 32 (4.7)LVI (%) 023 (53.5) 120 (46.5)PNI (%) 022 (51.2) 121 (48.8)Neoadjuvant therapy (%) 036 (83.7) 17 (16.3)TRG (%) 01 (14.3) 12 (28.6) 21 (14.3) 33 (42.9)Clinical and pathological features of esophageal cases. All cases were ESCC with R0 resection achieved. Pathological staging (pT and pN), LVI, PNI, neoadjuvant therapy status, and TRG for neoadjuvant-treated cases are shown. ESCC, Esophageal squamous cell carcinoma; LVI, lymphovascular invasion; PNI, perineural invasion; TRG, tumor regression grade. Clinicopathological characteristics for esophageal cases Clinical and pathological features of esophageal cases. All cases were ESCC with R0 resection achieved. Pathological staging (pT and pN), LVI, PNI, neoadjuvant therapy status, and TRG for neoadjuvant-treated cases are shown. ESCC, Esophageal squamous cell carcinoma; LVI, lymphovascular invasion; PNI, perineural invasion; TRG, tumor regression grade. FLI demonstrated a significant time advantage over conventional frozen-section analysis. Median imaging time was 5.4 minutes (IQR, 4.9-7.1) compared with 36.3 minutes (IQR, 31.2-40.3) for frozen-section preparation and interpretation (P < .001), as detailed in Figure E1. Model performance was assessed using 5-fold cross-validation. ROC curves from each fold for the lung tumor classification model are presented in Figure E2. The mean area under the curve across folds was 0.953. This time reduction was achieved by bypassing tissue sectioning and slide-based frozen-section workflows while maintaining robust predictive performance of the lung tumor classification model. Box-and-whisker dot plot comparing processing times for FLI imaging and frozen-section analysis. The lower and upper borders of each box indicate the first and third quartiles, respectively; the middle horizontal line represents the median; and the lower and upper whiskers represent the most extreme values within 1.5 times the IQR. Dots represent individual measurements. Values above the plot are presented as median time with IQR. FLI, Femtosecond label-free imaging; IQR, interquartile range. ROC curves of 5-fold cross-validation for lung tumor classification. Each curve represents 1-fold. ROC, Receiver operating characteristic; AUC, area under the curve. Morphological Concordance With Conventional Histopathology Across all specimens, independent side-by-side concordance review confirmed agreement between FLI and HE classification for all evaluated cases. Key diagnostic eatures, including malignant glandular architecture in lung adenocarcinoma, cellular pleomorphism, and keratin pearls in ESCC, and preserved architecture in benign/normal tissues were consistently identified on FLI images and found to correspond to the respective HE findings. Representative paired images are shown in Figure 1. FLI provided interpretable multimodal contrast across lung and esophageal tissues. In lung adenocarcinoma, tumor regions showed prominent THG signals with malignant architectural features, including irregular glandular structures, nuclear enlargement, and increased nuclear-cytoplasmic ratio, whereas adjacent normal lung tissue showed preserved alveolar architecture with relatively prominent 3PFE signals. In ESCC, tumor regions showed disorganized squamous architecture, cellular atypia, and intense THG signals corresponding to keratinizing areas, whereas adjacent normal esophageal mucosa retained orderly stratified squamous epithelium. These optical and morphological features were consistent with the corresponding HE findings. Depth Scanning for Margin Assessment FLI enabled assessment of surgical margins through sequential depth scanning without tissue sectioning. The margin-imaging workflow and 3-dimensional reconstruction process are shown in Figure 2, A. Gross inspection of the specimen provided only a surface-level assessment of the tumor-margin relationship (Figure 2, B). Likewise, superficial FLI imaging showed the tumor distribution at the tissue surface but could not fully represent tumor boundary changes at greater depths (Figure 2, C). Depth scanning-based margin assessment using FLI. A, Workflow of 3-dimensional reconstruction for margin specimens. B, Gross inspection of the lung adenocarcinoma specimen, illustrating the macroscopic appearance of the tumor at the resection margin. C, Superficial FLI image corresponding to the tissue surface, which, similar to gross inspection, is insufficient to assess variations in tumor boundaries across deeper tissue layers. D, FLI images obtained at increased imaging depths (D1: 27 μm; D2: 48 μm; D3: 69 μm) demonstrating depth-dependent tumor extent and morphology, with altered tumor–margin interfaces in deeper layers. White dashed lines indicate the changes in the boundary between tumor and normal regions across different depths. Three-dimensional reconstruction generated from sequential FLI depth images is provided in Video 2. 3D, 3-Dimensional; ROI, region of interest. Sequential FLI images were then acquired at 3-μm intervals from the tissue surface to a maximum imaging depth of 90 μm. Representative depth-resolved images demonstrated depth-dependent variation in tumor extent and morphology, with altered tumor-normal interfaces across different imaging planes (Figure 2, D). These findings illustrate that volumetric information within a sampled margin block may reveal spatial heterogeneity not captured by surface inspection or single-plane evaluation. Three-dimensional reconstruction based on sequential FLI depth images further illustrated the heterogeneous spatial distribution of tumor tissue relative to the surgical margin and is provided in Video 2. Exploratory Analysis for Prediction of High-Risk Pathological Features We further explored whether FLI could support intraoperative risk stratification in ESCC. We developed exploratory machine learning models to predict 3 clinically relevant high-risk pathological features: LNM, PNI, and LVI, using FLI-derived image features. Model performance was evaluated by 5-fold cross-validation. ROC curves from each fold are shown for LNM, PNI, and LVI (Figure 3, A-C). The mean area under the curve across folds was 0.653 for LNM, 0.766 for PNI, and 0.719 for LVI. Across the 5 folds, the models demonstrated consistent discriminative ability for all 3 end points, suggesting that FLI captures microstructural information associated with ESCC aggressiveness. These preliminary results suggest that FLI-derived image features may be associated with selected adverse pathological features in ESCC. However, this analysis was exploratory and was not intended to establish an intraoperative decision-making tool. Further validation is required to determine whether such predictions have clinical value. AI prediction of high-risk pathological features. A, ROC curves of 5-fold cross-validation for lymph node metastasis prediction. Each curve represents 1-fold. B, ROC curves of 5-fold cross-validation for perineural invasion prediction. C, ROC curves of 5-fold cross-validation for lymphovascular invasion prediction. ABMIL ROC, Attention-based multiple instance learning receiver operating curve; CV, cross-validation; AUC, area under the curve. Discussion In this prospective study of 355 thoracic surgical specimens, FLI provided rapid, slide-free imaging of fresh tissue with qualitative morphological concordance to paired histopathology and a shorter processing time than conventional frozen section. The principal findings are that FLI can be integrated into an intraoperative workflow, generate interpretable multimodal images without tissue sectioning or staining, provide depth-resolved information within sampled margin tissue, and support exploratory AI-based image analysis. Nonlinear optical microscopy techniques including SHG, THG, 2PFE, and 3PFE have established foundations in biomedical research.12 These techniques exploit endogenous tissue contrast: SHG reveals collagen architecture, THG highlights interfaces and lipid-rich structures, and multiphoton fluorescence captures metabolic substrates such as nicotinamide adenine dinucleotide and flavin adenine dinucleotide. These modalities enable label-free tissue visualization with subcellular resolution and optical sectioning capability, providing architectural information that parallels conventional histology.10,13,14 FLI technology integrates multiple nonlinear optical channels—THG, SHG, 2PFE, and 3PFE—into a unified platform driven by ultrashort femtosecond laser pulses. This multimodal architecture captures complementary tissue information in parallel, providing structural and compositional contrast without sectioning or staining. Recent work has validated this approach in breast tumor diagnosis,9 and the present study extends these findings to thoracic malignancies. The most immediate clinical advantage of FLI lies in supplementing or replacing frozen-section analysis. Conventional frozen-section pathology carries well-documented limitations. Turnaround time typically requires 20 to 40 minutes, during which surgical workflow is interrupted. Technical artifacts from ice crystal formation, tissue compression, and suboptimal sectioning can obscure diagnostic features.15,16 Sampling represents an inherent limitation, which means pathologists typically examine 1 or 2 tissue sections per specimen, risking missed focal pathology in heterogeneous tumors. FLI addresses these limitations through several mechanisms. The demonstrated median imaging time of 5.4 minutes represents a 4- to 7-fold reduction compared with frozen-section analysis, potentially decreasing anesthesia duration and improving operating room efficiency. By eliminating sectioning and staining, FLI avoids freezing artifacts and generates images with preserved cellular morphology. The AI interpretation framework offers standardized, reproducible assessments that reduce interobserver variability—a persistent challenge where diagnostic accuracy depends on individual expertise. This standardization could facilitate consistent care delivery across institutions. A notable illustrative case in our series involved a lung nodule in which FLI identified malignant glandular features (nuclear crowding, loss of polarity, and architectural distortion), whereas conventional HE showed cellular atypia with ambiguous architecture requiring thyroid transcription factor-1 immunohistochemistry for final confirmation (Figure E3). This case highlights a potential advantage of FLI in resolving diagnostically equivocal intraoperative findings, where conventional frozen section morphology alone may be insufficient. A challenging diagnostic case of lung adenocarcinoma: FLI clearly delineates dark yellow tumor cells and bright pink granulocytes in lung adenocarcinoma (scale bar: 100 μm). LUAD, Lung adenocarcinoma. The depth-scanning capability of FLI provides a distinct advantage for surgical margin assessment, particularly in thoracic procedures where margin adequacy determines resection extent and oncologic outcomes. Recent randomized trials have validated sublobe resection for early-stage lung cancer, but these limited resections demand rigorous margin evaluation.17,18 Conventional frozen-section evaluation examines only 1 or 2 tissue planes, potentially missing tumor involvement in adjacent but unsampled areas—a phenomenon termed “plane-selection bias,” reflecting the documented sampling-related discordance between frozen section and permanent pathology that has been reported as a key contributor to diagnostic error in thoracic oncology.2,3 This sampling limitation is particularly problematic in thick specimens where tumor distribution may vary across depth, as demonstrated in our study where some imaging showed limited margin involvement and other sections revealed more extensive tumor proximity. FLI overcomes this limitation through contiguous depth scanning at 3-μm intervals throughout tissue thickness (up to 90 μm), generating volumetric datasets that comprehensively represent specimen architecture. When indicated, these depth-series images enable 3-dimensional reconstruction, providing intuitive visualization of tumor-margin spatial relationships and reducing false-negative margin assessments caused by sampling bias. However, the clinical integration of volumetric FLI margin assessment remains exploratory, requiring prospective validation in controlled surgical cohorts to establish reproducible protocols and demonstrate impact on margin-positive rates and clinical outcomes. Beyond tumor detection and margin evaluation, we explored whether FLI-derived image features were associated with adverse pathological features in ESCC. This analysis should be interpreted cautiously. We agree that prediction of LNM, PNI, or LVI in ESCC is not expected to independently alter the extent of esophagectomy at the current stage. Therefore, the ESCC model was not intended to establish an immediately actionable intraoperative tool. Rather, it was included to test whether FLI-AI could detect image-based signals related to these pathological features and to define the current diagnostic capability and limitations of this approach. We also recognize that lung cancer high-risk features, such as STAS and visceral pleural invasion, may be more directly relevant to intraoperative decision-making because they can influence the adequacy of sublobar resection and the consideration of conversion to lobectomy in selected patients.19,20 However, the lung component of the present study was designed primarily for tumor classification, workflow feasibility, and morphological concordance, rather than for end point–specific modeling of STAS or visceral pleural invasion. Reliable analysis of these features would require targeted sampling, sufficient positive cases, and spatially matched pathological annotation between FLI images and corresponding histological findings. These requirements were not fully met in the current dataset. Accordingly, we have interpreted the ESCC high-risk feature analysis as hypothesis-generating and consider lung high-risk feature prediction an important direction for future FLI-AI studies. Limitations Current study limitations center on sample size and single-center design. Although our cohort provides proof-of-concept, the modest number of high-risk feature cases limits statistical power and generalizability of predictive models. The exploratory high-risk feature prediction models were developed and validated in a single-center cohort. This limited sample size constrains the statistical robustness and generalizability of these models. Thus, these results should be interpreted as proof-of-concept only. In addition, this study was primarily designed as a feasibility assessment of FLI for intraoperative pathological evaluation, focusing on diagnostic speed relative to conventional frozen section and diagnostic accuracy using permanent-section HE as the reference standard, rather than on comprehensive evaluation of all frozen-section–related pathological parameters. Ongoing efforts are expanding sample size in esophageal cancer to improve prediction accuracy, while establishing parallel models for lung cancer prognostic factors including STAS presence and lymph node metastasis risk. Multicenter validation will be essential to assess performance across different patient populations and institutional practices. FLI generates slide-free, label-free images morphologically concordant with conventional pathology in significantly reduced time. Its depth scanning capability addresses sampling limitations inherent to frozen-section analysis. Combined with AI-based image analysis, FLI may provide a promising framework for rapid and standardized intraoperative tissue assessment in thoracic oncology. However, the clinical role of AI-assisted margin evaluation and risk prediction remains exploratory and requires larger multicenter validation. Conclusions FLI provides rapid, slide-free intraoperative tissue assessment with qualitative morphological concordance to paired histopathology and a substantial workflow time advantage over conventional frozen section. Depth-resolved imaging offers additional within-block volumetric information for sampled margin tissue. Exploratory AI analyses suggest potential for future intraoperative adjunctive assessment, but these findings require validation before clinical implementation. Declaration of Generative AI and AI-Assisted Technologies in the Writing Process No use of any large language model in the research and manuscript development. Conflict of Interest Statement The authors reported no conflicts of interest. The Journal policy requires editors and reviewers to disclose conflicts of interest and to decline handling or reviewing manuscripts for which they may have a conflict of interest. The editors and reviewers of this article have no conflicts of interest. Footnotes This study was supported by the fellowship of China National Postdoctoral Program for Innovative Talents (No. BX20230083), National Natural Science Foundation of China (No. 82303564), and Youth Foundation of “Outstanding Resident Physician” Clinical Postdoctoral Program in Zhongshan Hospital (No. 2024ZYYS-031). Institutional Review Board approval: The equivalent ethics committee of the Zhongshan Hospital approved the study protocol and publication of data on May 16, 2025 (B2025-222R). The patients provided informed written consent for the publication of the study data. Data availability statement: The data that support the findings of this study are available from Dr Li (E-mail: li_m14@fudan.edu.cn). Drs Yin, Liu, Zhu, and Yu contributed equally to this article. Contributor Information Xin Zhu, Email: zhuxin@femtoresearch.com. Mingxiang Feng, Email: feng.mingxiang@zs-hospital.sh.cn. Lijie Tan, Email: tan.lijie@zs-hospital.sh.cn. Ming Li, Email: li_m14@fudan.edu.cn. Supplementary Data Video 1 FLI system operation and analysis workflow. Video available at: https://www.jtcvs.org/article/S2666-2736(26)00353-0/fulltext. Video 2 Three-dimensional reconstruction generated from sequential FLI depth imaging. Video available at: https://www.jtcvs.org/article/S2666-2736(26)00353-0/fulltext. Appendix E1 The entire operational process is presented in Video 1. (1) Specimen preparation (00:07-01:07) Fresh tissue samples were rinsed with normal saline and trimmed to approximately the size of a soybean. FLI specimens were collected in parallel with, but independent from, frozen-section specimens. The sample was placed in a confocal culture dish with careful removal of air bubbles between the sample and dish bottom at 00:07, and fixed with a holder at 00:18. A positioning image was subsequently captured at 00:32, followed by the selection of a region of interest for scanning at 00:41. (2) FLI scanning (01:08-04:25) The sample holder was loaded into the FLI system. After system initiation, automatic focusing was conducted between 01:16 and 02:04, and image acquisition was completed from 02:05 to 04:25. The number of specimens scanned per patient was determined according to the intraoperative sampling plan, including primary tumor, resection margin, and regional lymph nodes. (3) FLI-AI analysis (04:27-end) Scanning data were imported into the FLI-AI analysis system from 04:29 to 04:50. The system automatically identified tissue characteristics and differentiated benign and malignant tissues at 04:59, followed by the final output of diagnostic results. References Associated Data Supplementary Materials Video 1 FLI system operation and analysis workflow. Video available at: https://www.jtcvs.org/article/S2666-2736(26)00353-0/fulltext. Video 2 Three-dimensional reconstruction generated from sequential FLI depth imaging. Video available at: https://www.jtcvs.org/article/S2666-2736(26)00353-0/fulltext.