Researchers at Pusan National University in South Korea have developed a 3D printing-based diagnostic method that detects thyroid cancer using surface-enhanced Raman spectroscopy (SERS) on human serum. The technique bypasses the need for traditional biomarkers by analyzing spectral patterns with a convolutional neural network (CNN), achieving 93.1% sensitivity and 84.0% specificity. Results were published in Nature Communications.
Current diagnosis of thyroid cancer relies on fine-needle aspiration cytology, which can result in inconclusive outcomes and complications. Despite decades of research, no specific biomarker has been validated for the disease. The Pusan research team addressed this limitation by applying SERS to gold nanoparticle (AuNP) clusters fabricated directly from patient serum using an evaporation-based 3D printing process. These nanoclusters amplify Raman signals generated by biochemical compounds in the sample, allowing deep learning algorithms to classify cancer-positive and healthy cases without requiring a biomarker signal.
A total of 100 serum samples—50 from patients diagnosed with thyroid cancer and 50 from healthy individuals—were collected from the Human Derived Materials Bank of Pusan National University Yangsan Hospital. To fabricate the plasmonic substrate, researchers combined AuNPs (75 ± 5 nm diameter), saline, and human serum into a hybrid ink formulation. Saline was included to reduce albumin concentration, optimizing conditions for AuNP clustering. The ink was loaded into a 30 μm micropipette, and deposition was achieved by maintaining contact with a silicon substrate until solvent evaporation caused nanoparticle accumulation and cluster formation. Energy-dispersive X-ray spectroscopy confirmed the presence of dispersed AuNPs in the dried structures, while sodium chloride crystals were also observed due to saline content.
Each cluster formed localized electromagnetic hotspots through the interaction of AuNPs. These hotspots enhanced the Raman signal of nearby metabolites. The team recorded SERS spectra using a portable Raman spectrometer with a 633 nm laser at powers of 1.8 mW and 2.0 mW across four exposure durations: 500 ms, 1,000 ms, 2,000 ms, and 3,000 ms. A total of 800 spectra were collected, with 400 from each group. Measurements were taken across the Raman shift range of 200 to 1200 cm⁻¹, and spike noise was removed prior to analysis; no additional spectral filtering was applied. Repeatability was validated by performing five measurements per sample, confirming consistent spectral profiles within each individual’s serum.
To assess classification accuracy, both 1D and 2D CNN models were trained to distinguish between spectra from healthy individuals and cancer patients. Each spectrum was labeled as 0 or 1, and data were split 8:2 into training and test sets. The 2D CNN architecture included two convolutional layers (16 and 48 filters) and one dense layer with 256 neurons. A kernel size of 3×3 was used with a dropout rate of 0.6. The model employed the Adam optimizer, a learning rate of 0.0011949, and a sigmoid activation function. Training was conducted for 80 epochs. The resulting area under the curve (AUC) was 0.858, and the model demonstrated consistent performance across test samples.
Efforts to discover thyroid cancer biomarkers using nuclear magnetic resonance (NMR) analysis of 10 randomly selected serum samples (5 per group) failed to identify group-specific metabolites. Common compounds such as glucose, glucuronate, alanine, and glycerol appeared in both sets, reinforcing the absence of clear molecular indicators. This further justified the team’s decision to focus on spectral pattern recognition rather than biomarker detection.
Fabrication parameters were informed by earlier experiments using a hybrid ink composed of M13 bacteriophage and AuNPs. M13, a rod-shaped virus 880 nm in length and 6.6 nm in diameter, induced phase separation due to its geometric mismatch with spherical nanoparticles. Scanning electron microscopy revealed interspersed AuNP clusters and M13 aggregates, which increased the density of electromagnetic hotspots. However, the final serum-AuNP ink used in this study excluded M13 to simplify fabrication.
One limitation identified by the researchers was the relatively weak SERS signal in some samples. This was attributed to the low volume fraction of AuNPs compared to dominant serum proteins like albumin, which occupy more space within the printed structure. Increasing the nanoparticle concentration may improve spectral intensity but was not explored within the scope of the current study. Another constraint was the lack of an external validation cohort, which may limit the generalizability of the CNN model beyond the initial sample set.
Despite these constraints, the method presents a path forward for disease detection in cases where biomarker identification has stalled. The integration of nanostructured materials, spectroscopic readouts, and machine learning allowed for reproducible sample analysis using minimal input volumes and no invasive procedures. The authors note that similar methods may be adapted to other diseases with elusive molecular targets.
Want to help choose the 2025 3DPI Awards winners? Join the Expert Committee today.
Ready to discover who won the 2024 3D Printing Industry Awards?
Subscribe to the 3D Printing Industry newsletter to stay updated with the latest news and insights.
Featured image shows Nanoparticle-cluster for the SERS platform. Image via Nature Communication.
