Label-free pathological subtyping of non-small cell lung cancer using deep classification and virtual immunohistochemical staining

The differentiation between pathological subtypes of non-small cell lung cancer (NSCLC) is an essential step in guiding treatment options and prognosis. However, current clinical practice relies on multi-step staining and labelling processes that are time-intensive and costly, requiring highly specialised expertise. In this study, we propose a label-free methodology that facilitates autofluorescence imaging of unstained NSCLC samples and deep learning (DL) techniques to distinguish between non-cancerous tissue, adenocarcinoma (AC), squamous cell carcinoma (SqCC), and other subtypes (OS). We conducted DL-based classification and generated virtual immunohistochemical (IHC) stains, including thyroid transcription factor-1 (TTF-1) for AC and p40 for SqCC, and evaluated these methods using two types of autofluorescence imaging: intensity imaging and lifetime imaging. The results demonstrate the exceptional ability of this approach for NSCLC subtype differentiation, achieving an area under the curve above 0.981 and 0.996 for binary- and multi-class classification. Furthermore, this approach produces clinical-grade virtual IHC staining which was blind-evaluated by three experienced thoracic pathologists. Our label-free NSCLC subtyping approach enables rapid and accurate diagnosis without conventional tissue processing and staining. Both strategies can significantly accelerate diagnostic workflows and support efficient lung cancer diagnosis, without compromising clinical decision-making.

💡 Research Summary

This study introduces a fully label‑free workflow for pathological subtyping of non‑small cell lung cancer (NSCLC) that combines autofluorescence imaging (both intensity and fluorescence‑lifetime microscopy, FLIM) with deep learning (DL) classification and generative adversarial network (GAN)‑based virtual immunohistochemical (IHC) staining. Using tissue microarray (TMA) cores from over 280 patients (631 cores total), the authors captured two complementary imaging modalities: a single‑channel grayscale intensity image that reflects endogenous fluorophore brightness, and a four‑channel RGB FLIM image that encodes the decay lifetimes of key metabolic fluorophores such as NADH and FAD. Patches of 224 × 224 px were extracted (with overlapping patches for under‑represented classes) and fed into several state‑of‑the‑art convolutional neural networks (ResNet‑50, EfficientNet‑B3, DenseNet‑121, etc.).

Binary classification tasks (cancer vs. non‑cancer, adenocarcinoma (AC) vs. SqCC + other subtypes, SqCC vs. other, AC vs. SqCC) achieved area‑under‑the‑curve (AUC) values ranging from 0.94 to 0.999, with the FLIM‑based models consistently outperforming intensity‑only models, especially in the more challenging AC vs. SqCC discrimination (AUC > 0.97). Multi‑class (four‑class) classification of normal tissue, AC, SqCC, and other subtypes also yielded high performance, with AUCs between 0.981 and 0.996. Confusion matrices revealed that most errors occurred in distinguishing AC from SqCC, reflecting the known histopathological overlap.

Interpretability analyses using Grad‑CAM++ demonstrated that the FLIM‑based network focuses on metabolic signatures and stromal regions, whereas the intensity‑based network relies more on nuclear and cytoplasmic morphology. t‑SNE visualisation of the final fully‑connected layer embeddings showed well‑separated clusters for each class, with FLIM features providing slightly sharper boundaries, especially for the “other” (OS) category.

For virtual IHC, a conditional GAN previously used for H&E‑to‑IHC translation was retrained to synthesize thyroid transcription factor‑1 (TTF‑1) staining for AC and p40 staining for SqCC directly from the label‑free images. Quantitative image quality metrics (SSIM > 0.85, PSNR > 30 dB) indicated high fidelity, and a blind assessment by three experienced thoracic pathologists rated the synthetic stains as clinically comparable to real IHC.

The authors argue that (1) label‑free autofluorescence provides sufficient morphological and metabolic information for accurate NSCLC subtyping, (2) FLIM adds a valuable metabolic dimension that improves discrimination of histologically similar subtypes, and (3) GAN‑generated virtual IHC can replace or supplement conventional staining, preserving precious tissue for downstream molecular assays. Limitations include the need for specialized FLIM hardware, potential variability across imaging platforms, and the necessity of larger multi‑center validation before clinical deployment. Future work is suggested on expanding to additional lung cancer subtypes, integrating genomic/transcriptomic data, and developing portable FLIM devices to bring this technology into routine pathology labs. Overall, the paper demonstrates a promising, cost‑effective, and rapid alternative to traditional histopathology that could accelerate diagnostic workflows while maintaining, or even enhancing, diagnostic accuracy.