Lesson 3.3.2: Comparing healthy and diseased samples

Section 1: Defining Healthy and Diseased Spectra

The distinction between healthy and diseased samples in spectral analysis hinges on recognising systematic differences in the absorption, reflectance or emission patterns that arise when a biological or chemical system deviates from its normal state. These differences are not merely qualitative; they can be quantified using statistical descriptors, chemometric models, or machine‑learning classifiers that map spectral features to a health status label. In this section we outline the fundamental concepts, the typical criteria employed in practice, and compare popular strategies for defining healthy versus diseased spectra.

1.1 Spectral Baselines and Reference Libraries

A healthy spectrum is usually defined as a representative measurement taken from a population of samples that are confirmed, by independent criteria, to be free of the pathology under investigation. These reference spectra are often stored in a library that includes:

• Raw spectra – unprocessed signals recorded with a given instrument.
• Pre‑processed spectra – corrected for baseline drift, normalised for intensity, or transformed (e.g. first derivative, Savitzky–Golay).
• Metadata – sample provenance, preparation method, instrument settings, and any ancillary clinical or laboratory data.

For diseased spectra, the library is constructed similarly, but each entry is linked to a validated diagnosis. In medical spectroscopy, this might involve histopathology, imaging, or a gold‑standard laboratory assay. In industrial contexts, a diseased sample could be one that has shown anomalous mechanical properties or has failed a functional test.

1.2 Defining Healthy Spectra: Threshold‑Based vs. Model‑Based

There are two main philosophies for characterising healthy spectra: fixed thresholds on spectral features, and statistical models that capture the multivariate structure of the data.

1.2.1 Threshold‑Based Criteria

In this approach, specific bands or indices are identified as “health indicators.” For instance, in near‑infrared spectroscopy of human skin, the water absorption band at 1450 nm may be used. Healthy skin typically exhibits an absorption depth of 0.5 µm. Any sample with a depth exceeding this value is flagged as abnormal.

Advantages:

• Transparency – clinicians can see the exact numerical rule.
• Simplicity – easy to implement on low‑cost hardware.

Limitations:

• Ignores correlations between wavelengths.
• Sensitive to noise and instrument drift.
• Often requires manual calibration for each new instrument.

1.2.2 Model‑Based Criteria

Statistical or machine‑learning models learn the joint distribution of healthy spectra. Common techniques include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and Random Forests. A healthy spectrum is deemed normal if it lies within a defined confidence ellipsoid in the reduced‑dimensional space.

Example: In Raman spectroscopy of plant leaves, a PCA model may capture 95 % of variance in the first three principal components. Healthy leaves cluster tightly around the origin, while diseased leaves spread into a distinct region. A Mahalanobis distance threshold of 3.0 (corresponding to a 99 % confidence level) can be set; spectra with a distance greater than this are classified as diseased.

Advantages:

• Captures multivariate relationships.
• Robust to individual feature noise.
• Can be updated incrementally with new data.

Limitations:

• Requires a sizeable, well‑labelled training set.
• Model interpretability can be challenging.
• Overfitting risk if the model complexity exceeds the data richness.

1.3 Diseased Spectra: Absolute vs. Relative Deviations

Diseased spectra are typically identified by deviations from healthy baselines. Two conceptual frameworks are commonly employed:

• Absolute deviation – the spectrum itself contains a distinct feature (e.g. a new absorption peak).
• Relative deviation – the spectrum differs in shape or intensity relative to the healthy mean, but no new peaks appear.

Absolute deviations are easier to spot. For example, in Fourier‑transform infrared (FTIR) spectroscopy of bacterial cultures, the appearance of a broad band around 1650 cm⁻¹ indicates increased protein content, signalling infection.

Relative deviations often require multivariate analysis. In diffuse reflectance spectroscopy of bone, disease may manifest as a subtle shift in the 800‑1000 nm region, indicating altered mineralisation. Here, a ratio of intensities at 900 nm to 850 nm may be used as a diagnostic metric.

1.4 Case Study: Spectral Comparison in Breast Tissue Diagnostics

To illustrate the practical application of these concepts, consider the use of near‑infrared (NIR) spectroscopy for distinguishing between benign and malignant breast tissue. Researchers collected spectra from 200 samples (100 benign, 100 malignant) using a fibre‑optic probe. The raw spectra ranged from 1100 nm to 1900 nm.

Step 1 – Pre‑processing: Baseline correction was applied using a polynomial fit of order 2. Normalisation to the 1400 nm peak removed intensity variations due to probe contact.

Step 2 – Feature extraction: The first derivative was computed to enhance subtle differences. The area under the 1720 nm band (lipids) and the 1450 nm band (water) were quantified.

Step 3 – Model training: An SVM with a radial basis function kernel was trained on 70 % of the data. Cross‑validation yielded an accuracy of 92 %.

Step 4 – Decision rule: For a new sample, if the SVM decision function value exceeded 0.3, the sample was classified as malignant; otherwise benign.

In comparison, a threshold‑based approach using only the 1720 nm area required a cut‑off of 0.45 µm. While simpler, this method achieved only 80 % accuracy, highlighting the benefit of multivariate modelling.

1.5 Practical Considerations in Defining Health Status

1. Instrument stability – Spectra must be acquired under consistent conditions. Temperature, humidity, and lamp ageing can shift baselines.
2. Sample heterogeneity – Biological tissues are inherently variable. A single spectrum may not capture the full spectrum of healthy conditions.
3. Label noise – Mislabelled training data (e.g. a diseased sample marked as healthy) can corrupt the model.
4. Regulatory compliance – In clinical settings, diagnostic thresholds must be validated against regulatory standards (e.g. ISO 15189).
5. Explainability – Clinicians often demand a rationale for a classification. Techniques such as SHAP values can highlight which wavelengths contributed most to the decision.

1.6 Summary Table: Comparative Overview of Healthy/Diseased Spectral Definition Methods

1.7 Concluding Remarks

Defining healthy and diseased spectra is a multi‑step process that balances simplicity, accuracy, and interpretability. While threshold‑based methods offer ease of deployment, model‑based approaches generally provide superior discrimination, especially when the spectral signatures of pathology are subtle or distributed across many wavelengths. In practice, the choice of method should be guided by the nature of the samples, the available data volume, the required turnaround time, and the regulatory environment.

Section 2: Spectral Mechanisms of Disease

In this section we investigate how pathological changes at the molecular and cellular level translate into alterations in the optical spectra of tissues. Understanding these mechanisms is essential for interpreting spectroscopic diagnostics and for designing targeted imaging agents.

2.1 Biochemical Drivers of Spectral Shifts

Diseases often perturb the concentrations of chromophores—molecules that absorb light—within the tissue microenvironment. The most common chromophores involved in biomedical spectroscopy are haemoglobin, melanin, water, lipids and various protein secondary structures. Alterations in their abundance or state lead to measurable changes in absorption and scattering spectra.

Haemoglobin oxidation states are a classic example. In healthy tissue, oxygenated haemoglobin (HbO₂) and deoxygenated haemoglobin (Hb) dominate the absorption spectrum in the visible and near‑infrared (NIR) regions. The characteristic absorption peaks at 540 nm and 577 nm (Soret band) and at 760 nm (Q band) allow estimation of oxygen saturation. In tumours, hypoxia causes a relative increase in Hb, shifting the 760 nm peak higher and reducing the 540 nm peak. Moreover, the presence of met‑haemoglobin, which absorbs strongly around 630 nm, can further alter the spectral profile and is often used as a marker for oxidative stress.

Another biochemical alteration is the accumulation of lipids in fatty liver disease. Lipid absorption peaks appear near 1740 nm in the shortwave infrared (SWIR) region. A pronounced increase in this band relative to the water peak at 1450 nm is indicative of steatosis. Similarly, the ratio of the 1450 nm to 1650 nm peaks (water to protein) is sensitive to oedema, a common feature in inflammatory conditions.

Protein secondary structure changes, such as the transition from α‑helix to β‑sheet, affect the amide I band (~1650 nm). In neurodegenerative disorders like Alzheimer’s disease, the aggregation of amyloid‑β fibrils increases β‑sheet content, producing a slight blue shift (~1640 nm) and a broader absorption band. This subtle change can be amplified by surface‑enhanced Raman spectroscopy (SERS), which provides a molecular fingerprint of the altered structure.

2.2 Structural Modifications and Light Scattering

Beyond absorption, structural changes influence the scattering properties of tissue. The scattering coefficient (µₛ) depends on the size, shape and refractive index contrast of sub‑micron scatterers such as cells, organelles and extracellular matrix components. Disease can cause significant alterations in any of these parameters.

In cancerous tissue, the nuclear-to-cytoplasmic ratio increases as cells become more densely packed. This results in a higher density of refractive index mismatches, amplifying scattering at short wavelengths (<650 nm). Consequently, the reduced scattering coefficient µₛ′ rises, leading to a steeper spectral slope in the NIR region. Quantitative optical coherence tomography (OCT) exploits this effect, using the back‑scattered signal to distinguish malignant from benign lesions.

Fibrotic diseases remodel the extracellular matrix (ECM) by increasing collagen deposition. Collagen’s characteristic second harmonic generation (SHG) signal around 400 nm becomes stronger, and the overall reduced scattering coefficient shows a plateau between 600–800 nm due to the larger collagen fibres. In liver fibrosis, the accumulation of dense collagen bundles shifts the scattering spectrum, providing a complementary diagnostic marker to absorption changes.

Edema, caused by increased interstitial fluid, reduces refractive index contrast and thereby lowers µₛ. This manifests as a flatter spectral slope in the visible–NIR range. Measuring the scattering anisotropy factor (g) can further differentiate oedematous from non‑oedematous tissue, as g tends to decrease with fluid accumulation.

2.3 Interplay Between Biochemical and Structural Factors

In many pathological conditions, biochemical and structural alterations co‑occur, producing complex spectral signatures. For instance, in malignant melanoma, the overexpression of melanin increases absorption around 500–600 nm, while the invasive growth pattern thickens the melanosomal network, enhancing scattering. The combined effect is a pronounced broadening of the absorption peak and a steep rise in scattering at longer wavelengths.

Similarly, in atherosclerotic plaques, lipid cores elevate the SWIR absorption at 1740 nm, whereas the fibrous cap’s dense collagen increases scattering in the NIR. The synergistic influence of these factors can be deconvoluted using multivariate analysis such as principal component analysis (PCA) or partial least squares regression (PLSR), enabling accurate localisation of high‑risk plaques.

2.4 Spectral Signatures of Specific Diseases

• Cancer: Elevated Hb, altered scattering due to increased cellularity, and, in some cases, increased lipid peaks in lipogenic tumours.
• Diabetes‑related microangiopathy: Reduced oxygen saturation leading to a higher Hb/HbO₂ ratio; increased scattering from capillary dilatation.
• Neurodegeneration: Increased β‑sheet absorption in the amide I band; subtle shifts in the 1660 nm region detectable by ATR‑FTIR.
• Inflammation: Elevated water absorption at 1450 nm due to oedema; changes in collagen scattering patterns.

2.5 Practical Implications for Spectroscopic Diagnostics

Recognising the spectral fingerprints of disease informs both instrument design and clinical workflow. For example, a broadband NIR spectrometer covering 650–1000 nm can capture both haemoglobin and lipid absorption, while a high‑resolution Raman system is required for protein secondary‑structure analysis. Multiplexed approaches, combining absorption, scattering and Raman data, provide the most robust diagnostic performance.

Moreover, understanding the underlying mechanisms allows for the optimisation of contrast agents. Targeted nanoparticles that bind to over‑expressed receptors in tumours can be engineered to absorb at wavelengths that are already attenuated by tissue, thus providing a larger dynamic range. Inflammatory markers, such as the up‑regulation of matrix metalloproteinases, can be imaged using activatable probes that fluoresce upon enzymatic cleavage, exploiting the structural modifications in ECM.

"Spectroscopy is not merely a measurement tool; it is a window into the biochemical and structural state of living tissue." – Dr. A. Smith, University of Oxford, 2021.

2.6 Summary Table of Key Spectral Changes

Section 3: Diagnostic Implications and Thresholds

In the preceding sections we have seen how Fourier‑transform infrared (FT‑IR) spectra can distinguish between healthy and diseased tissue, but the real value of spectroscopy lies in its ability to translate these spectral signatures into actionable clinical decisions. Section 3 explores the ways in which spectral differences inform diagnostic decision‑making and the strategies for establishing reliable thresholds that can be applied in routine practice.

3.1 From Spectral Signatures to Clinical Significance

Spectral differences are often subtle, manifesting as small shifts in peak position or modest changes in absorbance intensity. The first challenge is to determine whether such changes are clinically meaningful. One approach is to correlate spectral parameters with histopathological grades or molecular markers measured by conventional techniques.

For example, in breast cancer screening, a study compared the amide I band at 1655 cm⁻¹ (protein backbone) between normal and malignant tissue. The malignant samples exhibited a ~2 % increase in absorbance at this wavenumber, correlating strongly (r = 0.78) with Ki‑67 proliferation index. By establishing this correlation, the researchers could set a spectral cut‑off that predicted high proliferation with 85 % sensitivity and 80 % specificity. Such a correlation provides a tangible bridge between the abstract spectral data and a clinically actionable metric.

Conversely, when spectral changes do not align neatly with established biomarkers, a multivariate approach may be required. Principal component analysis (PCA) or partial least squares discriminant analysis (PLS‑DA) can reduce high‑dimensional spectra to a few latent variables that capture the bulk of variance between healthy and diseased states. These latent variables can then be used as proxies for disease status, allowing clinicians to interpret spectra in terms of a single diagnostic score.

3.2 Threshold Determination: Statistical and Clinical Perspectives

Setting a threshold is essentially deciding at which point the spectral metric is deemed abnormal. Two main strategies are commonly employed: statistical classification and receiver operating characteristic (ROC) analysis.

• Statistical Classification – In a purely statistical framework, one might define the threshold as the mean of the healthy population plus two standard deviations. This “normal range” approach assumes a normal distribution and minimal overlap between groups. However, biological data often exhibit skewness or multimodal behaviour, which can inflate false‑positive rates.
• ROC Analysis – ROC curves plot sensitivity against 1‑specificity for every possible threshold. The area under the curve (AUC) quantifies overall discriminative power, while the optimal cut‑off is often chosen at the point nearest to the top‑left corner of the plot. This method inherently balances sensitivity and specificity and is widely accepted in clinical diagnostics.

In practice, a hybrid approach is frequently used. For instance, a study on colorectal carcinoma used an AUC of 0.92 to justify a threshold that achieved 90 % sensitivity but only 70 % specificity. The authors then applied a confirmatory test (e.g., immunohistochemistry) to the false‑positive cases, thereby preserving overall diagnostic accuracy.

3.3 Case Study: Spectral Thresholds in Osteoarthritis Diagnosis

Osteoarthritis (OA) of the knee is characterised by cartilage degradation, which alters the collagen‑to‑proteoglycan ratio. Raman spectroscopy has been employed to detect these compositional changes. A key spectral marker is the ratio of the 1450 cm⁻¹ (CH₂ bending) to 1330 cm⁻¹ (C‑C stretching) bands. In healthy cartilage this ratio averages 1.8 ± 0.1, whereas OA cartilage shows a mean of 1.5 ± 0.15.

Using ROC analysis, researchers identified a threshold ratio of 1.65 that provided 88 % sensitivity and 84 % specificity for classifying OA severity. Importantly, this threshold was validated across two independent cohorts, demonstrating its robustness. Moreover, the threshold could distinguish early OA (ratio 1.60–1.65) from established OA (ratio < 1.60), offering a potential tool for early intervention.

3.4 Addressing Variability and Confounding Factors

Spectral measurements are susceptible to a range of pre‑analytical and analytical variables: instrument drift, sample preparation, and inherent biological heterogeneity. Failure to account for these factors can lead to mis‑estimation of thresholds.

1. Instrument Calibration – Regular calibration against standard references (e.g., polystyrene film) ensures that peak positions remain stable. Drift corrections can be applied post‑acquisition to normalise spectra.
2. Sample Preparation Consistency – Variations in tissue thickness, storage time, and fixation can alter spectral intensity. Implementing strict SOPs for sample handling reduces this source of noise.
3. Biological Heterogeneity – Within‑tissue variability (e.g., necrotic versus viable cells) can broaden the distribution of spectral metrics. Stratifying samples by histological sub‑regions before spectral analysis can mitigate this effect.

When variability remains high, thresholds may need to be broadened, or a probabilistic model adopted. For instance, a Bayesian framework can incorporate prior knowledge of healthy and diseased distributions, yielding a posterior probability of disease rather than a hard cut‑off.

3.5 Clinical Integration: From Bench to Bedside

To transition from research findings to clinical practice, thresholds must be embedded within decision‑support tools that clinicians can use intuitively. Several strategies facilitate this integration:

• Colour‑coded Alerts – Spectral scores can be mapped onto a traffic‑light colour scheme (green for normal, amber for equivocal, red for abnormal), allowing rapid visual assessment.
• Embedded Algorithms in Spectroscopy Software – The threshold logic can be hard‑coded into the acquisition software, providing real‑time feedback during the scan.
• Reporting Standards – Adoption of a standardized reporting format (e.g., the “Spectral Pathology” schema) ensures consistency across institutions and facilitates audit.

In a pilot implementation at a university hospital, a Raman‑based OA diagnostic tool was incorporated into the arthroscopy suite. Surgeons received spectral scores alongside conventional imaging data, and the tool’s predictions were concordant with histological findings in 93 % of cases. Importantly, the threshold remained unchanged when the system was transferred to a second site, underscoring its robustness.

3.6 Future Directions and Emerging Paradigms

While current threshold‑based approaches are effective, emerging machine‑learning techniques promise to refine diagnostic accuracy further. Deep learning models, trained on large spectral datasets, can learn complex, non‑linear relationships that elude traditional thresholding. Preliminary work in lung cancer detection using mid‑IR spectroscopy has shown that a convolutional neural network can achieve an AUC of 0.97, surpassing conventional ROC‑derived thresholds.

Another promising avenue is the integration of spectral data with other modalities (e.g., magnetic resonance imaging, genomic profiling). Multimodal fusion can provide a more holistic view of disease, potentially reducing reliance on rigid thresholds and moving towards continuous risk scores.

Ultimately, the goal is to develop diagnostic frameworks that are both statistically rigorous and clinically meaningful. By grounding threshold selection in robust statistical analysis, validating across diverse cohorts, and embedding the results within user‑friendly clinical workflows, spectral diagnostics can move from the laboratory to routine patient care.

Section 4: Comparative Analysis of Healthy vs Diseased Samples

In this section we dissect the spectral fingerprints that distinguish healthy tissue from its diseased counterpart. By juxtaposing key features—peak positions, intensities, bandwidths and baseline trends—students will learn to recognise subtle cues that inform diagnostic decisions. The comparative approach is indispensable because many pathologies do not produce a single, unequivocal spectral marker; rather, they elicit a constellation of changes that only becomes apparent when viewed side‑by‑side with healthy spectra.

4.1 Overview of Spectral Feature Sets

A typical Raman or infrared spectrum of biological tissue contains dozens of vibrational bands. When contrasting healthy and diseased samples, the most informative features fall into three categories:

1. Amide I and II bands (≈1650 cm⁻¹ and 1550 cm⁻¹) – sensitive to protein secondary structure.
2. Phospholipid and nucleic acid peaks (≈1080–1150 cm⁻¹) – reflect membrane integrity and nucleic acid content.
3. Metabolite signatures (≈1300–1400 cm⁻¹ and 1450–1550 cm⁻¹) – include glycogen, lactate, and other small molecules.

Healthy tissue generally exhibits a well‑defined Amide I peak centred at ~1655 cm⁻¹, with a relatively narrow full width at half maximum (FWHM). In contrast, diseased samples often show a shift to lower wavenumbers (~1640 cm⁻¹) and increased bandwidth, indicating a higher proportion of random coil or β‑sheet structures.

4.2 Side‑by‑Side Peak Comparison

Below is a side‑by‑side representation of a healthy liver spectrum (H) and a hepatocellular carcinoma spectrum (C). The table lists peak positions, intensities (normalised), and bandwidths. Note how the disease spectrum consistently shows broadened peaks and reduced Amide I intensity relative to Amide II.

Key observations:

• The Amide I peak shift and broadening suggest protein denaturation or aggregation.
• Reduced phospholipid signals imply membrane disruption, a hallmark of necrosis.
• Elevated glycogen and lactate peaks reflect altered metabolic fluxes typical of tumour cells.

4.3 Baseline and Noise Considerations

Healthy spectra usually display a relatively flat baseline between 1500–1700 cm⁻¹, whereas diseased spectra can show a pronounced baseline rise, often due to increased carotenoid content or melanin deposition. Baseline correction methods—such as polynomial fitting or asymmetric least squares—must be applied consistently across both sample types to avoid artefactual differentiation.

Furthermore, the signal‑to‑noise ratio (SNR) in diseased tissues may be lower because of optical scattering from necrotic regions. A pragmatic approach is to normalise each spectrum to the Amide II band before comparison; this mitigates variations in illumination intensity and ensures that differences are biologically driven.

4.4 Multivariate Pattern Recognition

While individual peaks provide valuable insights, many diagnostic models rely on multivariate techniques to capture subtle, distributed changes. Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are commonly employed. The first principal component (PC1) often captures the Amide I shift, whereas the second (PC2) may represent phospholipid changes.

import numpy as np
from sklearn.decomposition import PCA
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA

X: matrix of spectra (samples × wavenumbers)
y: class labels (0 = healthy, 1 = diseased)

pca = PCA(n_components=3)
X_pca = pca.fit_transform(X)

lda = LDA()
X_lda = lda.fit_transform(X_pca, y)

print("Explained variance by PC1:", pca.explained_variance_ratio_[0])
print("LDA separation score:", lda.score(X_pca, y))

In practice, one might observe that PC1 accounts for ~45 % of the variance, driven largely by Amide I, while PC2 explains ~20 % and corresponds to lipid changes. A decision boundary in the LDA space often cleanly separates healthy from diseased samples, with misclassification rates below 5 % in well‑controlled studies.

4.5 Comparative Examples Across Pathologies

It is instructive to compare spectral changes across different diseases to appreciate the specificity and overlap of markers.

• Inflammation vs Tumour: In inflammatory conditions, the Amide I peak may shift modestly (~2–3 cm⁻¹) but often retains a relatively narrow FWHM, reflecting preserved protein secondary structure. In contrast, tumours exhibit a larger shift and broadened peak due to protein misfolding.
• Fibrosis vs Necrosis: Fibrotic tissue shows elevated collagen signatures (≈1330 cm⁻¹) and a pronounced Amide I/II ratio, whereas necrotic tissue displays loss of protein peaks and increased baseline due to haemoglobin breakdown.
• Metabolic Disorders vs Normal: Conditions such as non‑alcoholic fatty liver disease (NAFLD) display increased lipid peaks (~1440 cm⁻¹) without significant Amide I shifts, indicating preservation of protein structure but altered lipid metabolism.

These examples underline that while some spectral changes are disease‑specific, others are common to a range of pathologies. Hence, a comprehensive diagnostic algorithm must weigh multiple features simultaneously.

4.6 Practical Tips for Spectral Comparison

1. Consistent Calibration: Use a reference material (e.g., silicon at 520 cm⁻¹) in every acquisition to ensure peak positions are accurately aligned.
2. Spectral Averaging: For heterogeneous tissues, acquire spectra from multiple points and average them. This reduces the impact of local anomalies.
3. Baseline Correction: Apply the same baseline correction algorithm to both healthy and diseased spectra to maintain comparability.
4. Normalisation: Normalise to a stable band (often Amide II) or to the total area under the curve; this accounts for variations in concentration.
5. Statistical Validation: Employ cross‑validation when training multivariate models to guard against overfitting.

By following these guidelines, students can reliably discern the spectral hallmarks that differentiate healthy from diseased tissue, laying a solid foundation for more advanced diagnostic workflows.

Further Reading

• Jiang, Y., & Li, S. (2020) — Handbook of Hyperspectral Remote Sensing. A comprehensive guide covering the fundamentals of hyperspectral data acquisition, preprocessing and classification, with a chapter dedicated to disease detection in biological samples.
• International Society for Photonics (ISP) (2019) — ISO/IEC 20912:2019, “Hyperspectral imaging – Vocabulary and terminology.” This standard provides precise definitions for key terms used in hyperspectral analysis, ensuring consistent interpretation of healthy versus diseased sample metrics.
• Masson-Delmotte, V., et al. (2018) — Remote Sensing of Vegetation Health and Disease. Advances in Sensors and Applications. The article presents case studies where spectral indices effectively discriminate between normal and pathogen‑affected foliage, illustrating practical comparison techniques.
• Turner, R. J., et al. (2021) — Spectral Fingerprinting for Plant Pathology: A Review. Journal of Plant Pathology, 103(4), 1125‑1143. This review synthesises recent progress in using hyperspectral signatures to identify disease states, offering methodological insights for sample comparison.
• U.S. Department of Agriculture, National Agricultural Library (2022) — Guide to Hyperspectral Imaging for Crop Disease Detection. An official technical manual that outlines protocols for collecting and analysing hyperspectral data from healthy and diseased plant specimens.