Label‑Free Multiphoton Microscopy + ML for HCC: Accuracy and Tumor Border Maps

A Scientific Reports study describes a workflow that pairs label‑free multiphoton microscopy (MPM), using endogenous optical signals, with machine-learning classification to distinguish hepatocellular carcinoma (HCC) from non-neoplastic liver tissue—an approach the authors frame as potentially applicable to intraoperative assessment of resection planes.

The method relies on intrinsic contrast rather than exogenous labels and is presented as a way to analyze resection planes from native tissue signals. Alongside tissue classification, the authors report spatial outputs that visualize tumor borders from the same imaging data. Overall, the study centers on an image-based classifier paired with tumor–non-tumor border mapping derived from multimodal acquisitions.

For model development and testing, the authors report imaging matched samples of human HCC and background liver tissue from 76 patients. Imaging used three label-free multiphoton microscopy channels—coherent anti-Stokes Raman scattering (CARS), two-photon autofluorescence, and second harmonic generation (SHG)—with morphology from each channel reduced to 17 texture parameters used as model inputs. A supervised dataset was assembled at both patient and image levels: approximately 25,000 images from 35 patients were used for training, and a separate test set comprised approximately 27,000 images from 38 patients. The paper describes these texture features as the representation on which the classifier learned to separate neoplastic from non-neoplastic tissue patterns.

On the held-out test set, the authors report that the neural network achieved a test-set correct rate of 97.3% overall, with class-specific correct rates of 98.2% for liver and 96.5% for tumor; the abstract does not further define this metric (e.g., per-image vs per-patient) or describe clinical validation. These values are presented as the primary test-set performance summary for distinguishing HCC from non-neoplastic liver using texture parameters extracted from the multiphoton channels. Within this test set, the reported results indicate separation of tumor from background liver with a high correct rate.

Beyond image-level classification, the authors report creating maps showing the tumor border for three patients, describing these as visualizations that localize transitions between tumor and non-tumor predictions across imaged tissue regions. They also report that analysis of nonlinear-signal contributions to classification indicated autofluorescence played a key role in discriminating between neoplastic and non-neoplastic tissue. In addition, the paper states that accurate tumor recognition was achieved on low-lateral-resolution multiphoton images intended to mimic the use of endoscopes.

The abstract concludes that label-free intraoperative optical histopathology of HCC has the potential to improve tumor resection margins and, if implemented in endoscopes, may enable on-site tissue analysis.

Key Takeaways:

• The authors describe a multimodal, label-free multiphoton microscopy workflow incorporating CARS, two-photon autofluorescence, and SHG signals summarized as texture parameters for classification.
• High test-set discrimination between HCC and non-neoplastic liver tissue was reported using a supervised neural-network classifier trained on matched patient samples.
• Tumor border maps were reported for three patients, and the authors frame accurate recognition on low-lateral-resolution images as compatible with endoscopic-mimic implementation.