[Paper Review] Data Augmentation of Spectral Data for Convolutional Neural Network (CNN) Based Deep Chemometrics

Esben Jannik Bjerrum, Mads Glahder+1 more|arXiv (Cornell University)|Oct 5, 2017

Spectroscopy and Chemometric Analyses3 references136 citations

TL;DR

The paper shows that CNNs with data augmentation and extended multiplicative scatter correction (EMSC) outperform PLS models in predicting tablet drug content from NIR spectra, especially in extrapolation across instruments and concentrations.

ABSTRACT

Deep learning methods are used on spectroscopic data to predict drug content in tablets from near infrared (NIR) spectra. Using convolutional neural networks (CNNs), features are ex- tracted from the spectroscopic data. Extended multiplicative scatter correction (EMSC) and a novel spectral data augmentation method are benchmarked as preprocessing steps. The learned models perform better or on par with hypothetical optimal partial least squares (PLS) models for all combinations of preprocessing. Data augmentation with subsequent EMSC in combination gave the best results. The deep learning model CNNs also outperform the PLS models in an extrapolation chal- lenge created using data from a second instrument and from an analyte concentration not covered by the training data. Qualitative investigations of the CNNs kernel activations show their resemblance to wellknown data processing methods such as smoothing, slope/derivative, thresholds and spectral region selection.

Motivation & Objective

• Motivate the use of convolutional neural networks (CNNs) for regression on spectroscopic data to predict tablet drug content from NIR spectra.
• Evaluate preprocessing strategies (data augmentation and EMSC) against a baseline and hypothetical optimal PLS models.
• Assess model performance on data from two instruments and under extrapolation conditions.
• Investigate how CNN kernel activations relate to known spectroscopic processing concepts.

Proposed method

• Construct a CNN model in Keras/Theano with Gaussian noise, two 1D convolutional layers, dropout, and a linear output layer.
• Apply randomized data augmentation by perturbing offset, slope, and intensity to expand the training set.
• Compare preprocessing variants: no preprocessing, normalization, and Extended Multiplicative Scatter Correction (EMSC) with normalization.
• Optimize hyperparameters via Bayesian optimization (Gaussian processes) using a validation set; tune network architecture, dropout, and learning rates.
• Benchmark CNNs against Partial Least Squares (PLS) models using R2, RMSE, and Huber loss as metrics.
• Evaluate extrapolation by testing on data from a second instrument and higher assay values not present in training.

Experimental results

Research questions

• RQ1 Can CNNs learn to predict tablet assay concentration from NIR spectra as well as or better than PLS models across different preprocessing pipelines?
• RQ2 Does combining data augmentation with EMSC preprocessing yield the best predictive performance?
• RQ3 How do CNNs perform in extrapolation scenarios across instruments and unseen concentration ranges?
• RQ4 What do CNN kernel activations reveal about learned spectral features relative to traditional preprocessing (smoothing, derivatives, region selection)?

Key findings

• CNNs generally outperform PLS models on test sets across preprocessing variants when data augmentation is used.
• Data augmentation followed by EMSC provides the best performance among preprocessing strategies for CNNs.
• CNNs show strong extrapolation capabilities, outperforming PLS in an extrapolation challenge using a second instrument and unseen concentrations.
• EMSC improves CNN performance consistently, while augmentation alone often helps more for CNNs than for PLS.
• Qualitative kernel activation analysis suggests CNNs learn features resembling smoothing, derivative/slope recognition, and spectral region selection, indicating interpretable behavior.
• The extrapolation test highlights the importance of hyperparameter optimization aligned with intended future performance.