The project's technical work was structured around four main work packages, which encountered significant methodological challenges that led to strategic pivots and important new developments.
The project's most significant methodological challenge was the preparation of formalin-fixed and paraffin-embedded (FFPE) human tissue samples for multimodal analysis.
Work on LIBS (Laser-Induced Breakdown Spectroscopy) preparation (WP1) determined that direct analysis on bulk FFPE blocks was unfeasible due to high spectral interferences from the paraffin matrix (e.g. C2 and CN bands). Analysis on thin sections was successful, but the substrates required for MALDI-MSI (glass or conductive ITO-coated slides) produced high background signals that saturated the LIBS detectors. This critical finding indicated that a co-located measurement on the exact same sample was not viable
This use of the kHz-LIBS system required a substantial adaptation of the preceding kHz-LIBS system to make it suitable for biological analysis. The previous setup was designed for geology and used a laser distance sensor for focusing, which failed to respond to translucent tissue on transparent slides. The main technical achievements were:
1. New Focusing System
2. Dual-Detection System: A critical challenge was "dynamic range saturation," where major elements saturated the detectors, hiding the trace elements of interest. This was solved by re-engineering the system to include two simultaneous detection paths.
3. Hardware Integration: A new sample holder was developed, and the entire system was precisely synchronized.
The outcome was a "functional kHz system" that successfully detected inorganic Ti nanoparticles in the rodent tissue models. A strategic deviation was made prioritize spatial resolution (10 µm), a 2x improvement on the original project’s goal
This Data Processing and Fusion Workflow became the principal focus of the project's efforts in order to build a data treatment pipeline.
The initial "low-level" fusion strategy (concatenating datasets) failed. The corrective action was to pivot to a "mid-level" (or "feature-level") fusion strategy. This involved developing discrete, functional scripts for preprocessing (e.g. 'Rolling Ball' background removal, 'cross-correlation' image alignment). This workflow statistically correlates 2D elemental maps from LIBS with thousands of individual molecular (m/z) signals from MALDI.
This work was supported by two key collaborations that produced new data science methods:
Hyperspectral Denoising: An investigation into denoising techniques (e.g. Savitzky-Golay, FFT, PCA, Whittaker Smoothing). The key discovery was that PCA is the most effective method for this LIBS framework, providing a 5-fold SNR enhancement with no signal distortion. This was the first known application of PCA and Whittaker Smoothing to LIBS data in this context. The Automated Peak Identification was developed to automatically identify emission lines in LIBS spectra.
