What question did this study set out to answer?

The aim is to develop a rapid, accurate intraoperative diagnostic tool for lung cancer using a novel imaging and AI platform.

May 29, 2026

View Full Paper

Rapid multi-task intraoperative diagnosis of lung cancer via deep neural network-driven label-free femtosecond laser imaging (FLI).

YLYin LiShanghai Jiao Tong University YWYue WangUniversity of Stuttgart TPTao PanShanghai Jiao Tong University

Key Points

The aim is to develop a rapid, accurate intraoperative diagnostic tool for lung cancer using a novel imaging and AI platform.
Developed an integrated FLI+AI platform named FastLung for multi-task diagnosis.
Utilized 326 paired tumor-normal specimens from surgical resections with multimodal FLI imaging.
Employed a self-supervised deep learning model trained on ~4 million image patches for real-time assessment.
FastLung achieved AUC = 0.9854 for benign vs. malignant assessment, AUC = 0.9573 for invasive vs. minimally invasive adenocarcinoma, and AUC = 0.9892 for adenocarcinoma vs. squamous cell carcinoma.
Reduced diagnostic turnaround time significantly to 5-10 minutes compared to 20-30 minutes for traditional frozen sections (60-80% faster).
Achieved an accuracy of 95.8% in core needle biopsies, with consistent performance across operators and time.

Abstract

8019 Background: Rapid and accurate intraoperative pathological diagnosis is critical for guiding surgical margins and preserving lung function during lung cancer surgery. Current frozen-section (FS) analysis is time-consuming and prone to artifacts. Femtosecond label-free imaging (FLI) enables high-resolution, non-destructive tissue visualization without staining or freezing, offering a promising platform for real-time assessment. Methods: We developed FastLung, an integrated FLI+AI platform for multi-task diagnosis. Fresh, unprocessed tissue samples (326 paired tumor-normal specimens) from surgical resections were imaged using multimodal FLI. FastLung employs a self-supervised deep learning model trained on ~4 million image patches, with formalin-fixed paraffin-embedded (FFPE) histopathology as ground truth. The system provides malignancy probability maps and categorical outputs within 5 minutes. Results: FastLung achieved high diagnostic performance across key intraoperative tasks: benign vs. malignant (AUC = 0.9854 ± 0.01), invasive vs. minimally invasive adenocarcinoma (AUC = 0.9573 ± 0.03), and adenocarcinoma vs. squamous cell carcinoma (AUC = 0.9892 ± 0.009). It significantly reduced diagnostic turnaround time (5–10 min vs. 20–30 min for FS, 60–80% faster) while maintaining accuracy comparable to or exceeding FS. In core needle biopsies (n = 94), accuracy reached 95.8%. The platform also demonstrated consistent performance across operators and time, supporting reliable integration into surgical workflow. Conclusions: FastLung combines label-free FLI with deep learning to deliver rapid, accurate, and reproducible intraoperative diagnosis of lung cancer, outperforming frozen sections in speed and multidimensional assessment. Its robust performance in both resection and biopsy specimens highlights its potential to improve surgical decision-making and extend toward pan-cancer diagnostic applications.

AIに質問

Bookmark

View Full Paper

AIに質問

Bookmark

View Full Paper

Rapid multi-task intraoperative diagnosis of lung cancer via deep neural network-driven label-free femtosecond laser imaging (FLI).

Key Points

Abstract

Cite This Study

Also Consider

Also Consider