Science & Methodology

The science behind ZinMed

How Raman spectroscopy and AI are used to support skin cancer screening — principle, method, explainability, data and publications. Presented honestly for the research stage.

Investigational · VNU project QG.24.84 · Device not yet licensed. This is a screening-support tool; the result is an AI-estimated risk indicator — not a diagnosis and not a substitute for a physician. The device is under research.
Principle

Raman spectroscopy — the molecular “fingerprint” of tissue

When a 785nm laser illuminates skin tissue, most scattered light keeps its wavelength; a tiny fraction shifts with molecular bond vibrations (Raman scattering). The resulting spectrum is a biochemical fingerprint of the tissue.

Biochemical changes from cancer cells (proteins, lipids, DNA…) alter intensity at certain wavenumbers. This is the signal the AI learns to estimate a risk indicator — non-invasive, no tissue sampling.

4001000180026003400cường độ (a.u.)

Illustrative Raman spectrum (not patient data). Orange lines: the 8 wavenumbers the AI relies on most (XAI).

AI method

From raw spectrum to risk indicator

A signal-processing + machine-learning pipeline, designed to be lightweight and run on-device (offline).

1

Pre-processing

Savitzky–Golay smoothing + SNV normalization to remove noise and baseline drift.

2

Dimensionality

PCA (50 components) compresses the 1501-point spectrum, keeping most variance.

3

Classification

Random Forest (400 trees) into 4 classes: BCC · SCC · Melanoma · Benign.

4

Deployment

Exported to ONNX (~3.4 MB) for lightweight, offline on-device inference.

Explainable AI (XAI)

8 wavenumbers the model relies on — interpretable

Instead of a black box, ZinMed traces the wavenumbers (cm⁻¹) the model attends to most, which map to known biochemical bonds. This is a core differentiator versus giving only a number.

856
cm⁻¹
990
cm⁻¹
1104
cm⁻¹
1358
cm⁻¹
1420
cm⁻¹
1602
cm⁻¹
1748
cm⁻¹
2960
cm⁻¹
Data & Targets

Model-development dataset

The model-development set contains ~1,200 Raman spectra — ~500 measured from clinical samples and ~700 simulated — used for internal training and testing. Composition is as declared by the research group; not independently validated.

Design target: 75–85% sensitivity and specificity. This is a research goal, not a clinically validated result.

Safety & Compliance

Designed to standards (under evaluation)

Designed toward Class 1 laser safety (IEC 60825) and medical electrical safety (IEC 60601) — under evaluation.

Building a quality-management system aligned to ISO 13485 (not yet certified).

Publications

A foundation of peer-reviewed work

The team has a substantial publication record in Raman + AI for biomedical diagnosis (diabetes, cancer, devices). A selection below.

A Hybrid 1D-CNN and Transformer Architecture for Differentiating Malignant Melanoma from Non-Melanoma Skin Cancers using Raman SpectroscopyRaman · ung thư da
Non-invasive in vivo Type 2 Diabetes Mellitus diagnosis using Raman spectroscopy in combination with Machine LearningMobile Networks and Applications
Machine learning approach for early detection of diabetes using Raman SpectroscopyMobile Networks and Applications 29:294–305
Prediction of the change of human blood glucose from Raman scattering by polynomial data preprocessingEAI ICCASA
A practical approach for colorectal cancer diagnosis based on machine learningAI chẩn đoán
Predicting the level of hypertension using machine learningAI chẩn đoán
Detection and classification of knee osteoarthritis using YOLOv3 and VGG-16 modelsAI chẩn đoán hình ảnh
Microfluidic impedance biosensors for cancer detection & monitoringThiết bị y sinh
Research on smart mobile application and cloud technology for monitoring and diagnosing diabetesThiết bị & phần mềm y tế

Research collaboration & clinical evaluation

Are you a clinician, researcher or health facility wanting to help validate this technology? Get in touch.

← Home