A 5-minute laser scan beat the 36-minute frozen section in a Shanghai lung-cancer OR trial

A Shanghai thoracic surgery team is piloting a different kind of intraoperative microscope. Instead of slicing a fresh lung-tumor specimen, freezing it, staining it, and reading it under a conventional slide, they point an ultrashort-pulse laser at the tissue and let a deep-learning model classify what comes back. The result, in a 144-patient prospective study at Shanghai Zhongshan Hospital, is a median 5.4-minute read on fresh lung-tumor specimens versus 36.3 minutes for the standard frozen-section pathology workflow.

The technique is called femtosecond label-free imaging, and it works on a different physical principle. The laser fires pulses about a quadrillionth of a second long, so short that the tissue's own molecules respond with several nonlinear optical signals at once: third harmonic generation (light at three times the laser's frequency, picked up from lipid-rich structures), second harmonic generation (light at twice the frequency, sensitive to ordered collagen), and 2-photon and 3-photon fluorescence (intrinsic signals from metabolic cofactors and proteins). No dyes are needed because the contrast is built into the tissue chemistry. No sectioning is needed because the pulses penetrate the specimen; the scanner reads depth at 3-micrometer intervals down to 90 micrometers, capturing volumetric margin information a single-plane frozen section cannot give a surgeon.

The published numbers come from a study in JTCVS Open (PMID 42604329), covering 144 patients enrolled between June and December 2025. Researchers collected 259 fresh lung specimens and 96 fresh esophageal specimens. For lung-tumor classification, the team's pipeline combined the UNI v1 pretrained histology foundation model with an attention-based multiple-instance learning head. Mean AUC reached 0.953, with specimen-level and patient-level splits used to prevent the data-leakage problem that has inflated many AI pathology results. Median turnaround was 5.4 minutes for the femtosecond scan versus 36.3 minutes for frozen-section pathology (P < .001).

The esophageal arm of the study is exploratory. Predicting high-risk features such as lymph-node metastasis and perineural invasion from the same optical data hit mean AUCs of 0.653 to 0.766, modest enough that the team is right to label the lung results as the headline and the esophageal numbers as a signal worth chasing in a tumor type where frozen-section margins are notoriously hard.

The result is not a replacement for frozen-section pathology. It is evidence that intraoperative pathology is shifting off the section-stain-microscope stack and onto a nonlinear-optics-plus-foundation-model stack. The unit of analysis changes from a stained two-dimensional slide to a volumetric optical volume, and the read is done by a model pretrained on a different imaging modality entirely. The throughput is different: a cryostat and a pathologist handle one specimen at a time, while the optical scanner finishes in single-digit minutes.

Two things need to happen before any thoracic surgery workflow changes. First, the AUC 0.953 and the time gain need to reproduce outside Shanghai Zhongshan Hospital, on different scanners, with different surgeons and a different patient mix. Second, and more important, the study did not report a clinical-outcome endpoint: re-excision rate, time to adjuvant therapy, operating-room turnover, or cost. A faster read is a workflow gain. Whether it changes what happens to the patient is a separate question, and one this study does not answer.

Frozen-section pathology remains the standard of care. What the Shanghai result establishes is that the alternative stack is now plausible enough to test in a multi-center trial with patient outcomes as the endpoint. The next paper to watch is a multi-site read on whether women and men on the operating table get home sooner, or with cleaner margins.