Ewing sarcoma (ES) is the second most common primary bone malignancy in children and young adults, occurring at a rate of 2.9 cases per million population in those under 20 years of age. It originates from primitive mesenchymal stem cells and presents as a poorly differentiated, highly aggressive tumor of bone or soft tissue. With modern multimodal therapy combining chemotherapy, surgery, and radiotherapy, overall survival rates have improved substantially, but early and accurate diagnosis remains the single most critical determinant of outcome.
The osteomyelitis problem: Osteomyelitis is a bacterial infection of bone, most commonly caused by Staphylococcus aureus and predominantly haematogenous in origin in pediatric patients. It shares a remarkably similar clinical picture with Ewing sarcoma: both conditions present with localized bone pain, swelling, fever, and elevated inflammatory markers. On plain radiographs and even MRI, the two conditions can produce nearly indistinguishable imaging findings, including periosteal reaction, cortical erosion, and marrow edema.
Why misdiagnosis is so costly: When osteomyelitis is mistaken for Ewing sarcoma, patients may receive unnecessary aggressive chemotherapy and surgery with significant morbidity. When Ewing sarcoma is misdiagnosed as osteomyelitis, patients lose critical weeks or months during which the tumor continues to grow and potentially metastasize. Documented cases exist in the literature where ES was treated as chronic osteomyelitis for months before biopsy revealed the true diagnosis, at which point the window for optimal treatment had narrowed considerably.
The current diagnostic pathway: The gold-standard diagnostic workup for suspected Ewing sarcoma involves open biopsy followed by histopathological examination combined with immunohistochemistry and cytogenetic analysis for the characteristic EWSR1 gene translocation. This process is invasive, time-consuming, and not without risk. Needle biopsy is increasingly used but carries a diagnostic yield that varies with operator experience and tumor heterogeneity. Both approaches require institutional expertise that is not universally available, creating a diagnostic access gap in lower-resource settings.
Fourier Transform Infrared (FTIR) spectroscopy is a technique that measures how different molecular bonds within a biological sample absorb infrared radiation. Every class of biomolecule, including nucleic acids, proteins, phospholipids, and polysaccharides, absorbs infrared light at characteristic wavenumbers (measured in cm-1), producing a unique spectral signature. When tissue undergoes pathological change, the relative concentrations of these biomolecules shift, and those shifts are encoded in the infrared spectrum.
Attenuated Total Reflection (ATR) mode: This study used ATR-FTIR, where the tissue specimen is pressed directly against a diamond crystal and the infrared beam undergoes total internal reflection within it, creating an evanescent wave that penetrates only a few micrometers into the sample surface. This mode requires minimal sample preparation, produces highly reproducible spectra, and avoids water absorption interference that complicates transmission FTIR in biological samples. Spectra were recorded on a Bruker Vertex 70v spectrometer across 32 scans at 2 cm-1 spectral resolution.
Why bone tissue specifically: Bone is a composite of collagen-based protein matrix and hydroxyapatite mineral. The protein component, primarily type I collagen, contributes strong amide I (approximately 1650 cm-1) and amide II (approximately 1550 cm-1) bands. The mineral component contributes phosphate bands around 1000-1100 cm-1. Tumor infiltration and bacterial infection each alter this biochemical architecture in distinct ways, raising the hypothesis that FTIR spectra might encode diagnostically useful information that visual inspection alone cannot capture.
Prior evidence in cancer diagnostics: FTIR spectroscopy had already been validated for distinguishing malignant from benign tissue in melanoma, cervical cancer, breast cancer, gastric cancer, and brain tumors before this study. Its application to bone sarcomas versus infection, however, remained largely unexplored, with only a preliminary report on Ewing sarcoma specifically, making this study a meaningful extension of the field.
This retrospective study enrolled 27 patients with confirmed Ewing sarcoma (ages 5-20 years, median age 14, male/female ratio 12:15) and 10 patients with confirmed osteomyelitis (ages 2-17 years, median age 11, male/female ratio 6:4). Samples of bone tissue were obtained during diagnostic or therapeutic procedures. Three distinct tissue groups were analysed: normal bone tissue sampled from outside the ES infiltration zone (serving as an internal control), ES-infiltrated bone tissue, and osteomyelitis-affected bone tissue. This three-group design allowed the authors to anchor pathological spectra against a normal baseline from within the same patient cohort.
Sample preparation protocol: Tissue specimens underwent a standardized fixation and dehydration protocol before FTIR measurement. Samples were placed in liquid fixative for approximately 12 hours, then underwent a graded ethanol series (50%, 70%, 80%, 90%, 96%) before full xylene infiltration and paraffin embedding. This is a conventional histopathology preparation method, meaning FTIR analysis could in principle be performed on the same paraffin blocks already created during routine diagnostic workup, without requiring additional tissue.
Amide I deconvolution: A critical analytical step was the curve-fitting deconvolution of the amide I spectral region (1600-1700 cm-1) to resolve overlapping protein secondary structure components. Second-derivative spectra were used to locate initial peak positions, and Gaussian functions were fitted to individual sub-peaks using MagicPlot 2.7.2 software. This allowed quantification of the alpha-helix and beta-sheet protein secondary structure fractions within each tissue type, providing a window into how the protein matrix changes with disease.
Statistical and machine learning pipeline: Raw spectra underwent vector normalization and baseline correction using OPUS 7.0 software. Statistical significance between groups was assessed by one-way ANOVA followed by Tukey's post-hoc test, with p less than 0.05 as the significance threshold at a 95% confidence interval. Machine learning classification was then applied using Python's scikit-learn library, encompassing both unsupervised dimensionality reduction and supervised classifier training and validation against an external test set.
The normalized average FTIR spectra of all three tissue groups showed clear, reproducible absorption peaks attributable to nucleic acids, phospholipids, polysaccharides, proteins, and lipids. While all three groups shared the same general spectral architecture, peak positions and maximum absorbance values differed systematically between groups. Peak shifts were observed between osteomyelitis and ES relative to the normal bone control, and also between ES and osteomyelitis with each other, confirming that disease-specific biochemical remodeling produces measurable spectral divergence.
DNA and phospholipid bands: The most prominent spectral differences between osteomyelitis and ES were found in the wavenumber regions corresponding to DNA, phospholipids, amide II, and lipids. The phosphate band from DNA, RNA, and phospholipids, centered near 1029 cm-1 in normal tissue, shifted to 1041 cm-1 in osteomyelitis and further to 1059 cm-1 in ES tissue. Osteomyelitis showed the highest maximum absorbance values in this region, reflecting altered nucleic acid and membrane phospholipid metabolism driven by the inflammatory response and bacterial infection. Lipid peroxidation, a known consequence of inflammation, likely contributes to altered lipid-band profiles in osteomyelitis.
Protein secondary structure via amide I deconvolution: Deconvolution of the amide I region revealed that the three tissue types differ not only in total protein content but in the ratio of alpha-helix to beta-sheet secondary structure. Normal bone outside the ES zone showed a mean alpha-helix to beta-sheet ratio of 0.80 (plus or minus 0.02), an alpha-helix fraction of 44.4% and a beta-sheet fraction of 55.6%. In ES tissue, absolute absorbance values for both alpha-helix (0.15 plus or minus 0.03) and beta-sheet (0.21 plus or minus 0.07) were significantly reduced compared to normal tissue (p less than 0.05), reflecting the displacement of the native protein matrix by poorly differentiated tumor cells with a distinct proteome. Osteomyelitis tissue showed ten resolvable band components in the amide I region compared to seven in normal bone and eight in ES tissue, suggesting greater protein structural heterogeneity from the inflammatory infiltrate.
Lipid profiles and disease-specific changes: The differences in the lipid spectral region between ES and osteomyelitis are consistent with distinct underlying pathophysiology. ES, as a rapidly proliferating tumor, shows altered membrane lipid composition and turnover. Osteomyelitis, driven by bacterial infection, triggers substantial lipid peroxidation and membrane disruption in host tissue. These distinct mechanisms produce measurably different lipid infrared signatures, adding a second dimension of spectral discrimination beyond the protein and nucleic acid bands.
With raw FTIR spectra spanning hundreds of wavenumber variables per sample, dimensionality reduction is necessary to visualize whether tissue classes naturally cluster in lower-dimensional space. The authors applied a broad suite of both linear and non-linear (manifold) dimensionality reduction methods, including Principal Components Analysis (PCA), Factor Analysis, Fast Independent Components Analysis (FastICA), Incremental PCA, Truncated Singular Value Decomposition (SVD), and Kernel PCA with both linear and radial basis function (RBF) kernels.
PCA results and limitations: Standard PCA, which finds linear combinations of spectral variables that explain the most variance, produced poor separation between the three tissue classes in the first two principal components. Particularly, the normal bone tissue and ES tumor tissue clusters showed substantial overlap in two-dimensional PCA space. This indicates that the most obvious sources of spectral variation are not the disease-discriminating features, and that the diagnostic signal is distributed across higher-order components not captured by simple two-component PCA plots.
Non-linear methods and higher-order components: The non-linear manifold learning methods, including Kernel PCA with an RBF kernel, offered improved visual separation compared to linear PCA, consistent with the idea that the disease-relevant spectral variation is non-linearly distributed across wavenumber space. The authors further reasoned that higher-order PCA components, not just the first two, might carry discriminating power, which motivated the transition to supervised classification methods that can exploit all available components simultaneously rather than relying on the first two principal components for visualization.
The unsupervised analysis served an important exploratory function: it confirmed that spectral differences between the three tissue classes are genuine but subtle in the most prominent variance directions. This is actually a common finding in tissue FTIR studies, where the biological variation within disease classes can rival the variation between them in overall spectral space. Supervised learning, which is trained with class labels to find the most discriminating features rather than the most variable features, is therefore the appropriate next step.
After dimensionality reduction, a range of supervised machine learning classifiers were trained on the reduced spectral data using class labels (normal, ES, osteomyelitis). The classifiers were evaluated against an external test set, a held-out portion of the data not used during training, providing a more realistic estimate of generalization performance than simple cross-validation on the training set. The best-performing classifier was identified as Quadratic Discriminant Analysis (QDA).
QDA mechanics: QDA is a generative probabilistic classifier that models each class as a multivariate Gaussian distribution with its own class-specific covariance matrix. Unlike Linear Discriminant Analysis (LDA), which assumes equal covariance across classes and produces linear decision boundaries, QDA fits a separate covariance matrix per class, allowing quadratic (curved) decision boundaries. This flexibility is appropriate when the spectral feature distributions of the three tissue types differ not just in mean but in their spread and correlation structure, which is biologically plausible given the distinct pathological mechanisms.
Classification performance on the external test set: The QDA classifier achieved an overall accuracy of 80% on the external test set. Per-class precision, recall, and F1-score values were as follows: for normal bone tissue, precision 0.78, recall 0.78, F1-score 0.78; for Ewing sarcoma tissue, precision 0.75, recall 0.86, F1-score 0.80; for osteomyelitis tissue, precision 0.88, recall 0.78, F1-score 0.82. Notably, ES recall was the highest of the three classes at 0.86, meaning the classifier correctly identified 86% of ES samples, an important property for a cancer-screening-adjacent application where missing a true positive is costly.
Gradient boosted classifier for pairwise discrimination: For the specific ES-versus-osteomyelitis two-class problem, a gradient boosted classifier was applied and achieved high accuracy in separating the two key diagnostic entities from one another. Gradient boosting is an ensemble method that builds a strong classifier by iteratively training shallow decision trees, each one correcting the residual errors of its predecessors, and combining their weighted predictions. The ability of a gradient boosted tree ensemble to handle non-linear feature interactions with limited data makes it well-suited for small-cohort spectral datasets like this one.
The significance of this study lies not in replacing histopathology, but in establishing whether FTIR spectroscopy can serve as a rapid, objective, and adjunct diagnostic signal in the evaluation of bone lesions where ES and osteomyelitis cannot be distinguished by clinical or radiological means alone. The authors explicitly position FTIR alongside, rather than in competition with, routine radiological and histopathological methods. Given that FTIR can be performed on standard paraffin-embedded tissue blocks that are already produced during any bone biopsy, the incremental cost and sample demand of adding FTIR analysis are low.
Speed advantage: Histopathological review of bone biopsies requires decalcification, a process that can take several days, followed by sectioning, staining, and expert pathologist review. Immunohistochemistry and cytogenetics for EWSR1 translocation detection add further days. FTIR spectral acquisition is completed in minutes per sample with near-immediate computational analysis. In a clinical pathway where diagnostic delay directly worsens outcomes, even a rapid preliminary FTIR classification could influence the urgency with which molecular confirmation is pursued.
Addressing the biochemical basis of misdiagnosis: The histological similarity between ES and osteomyelitis is well-documented. Both diseases produce small round cells in the bone marrow space, and the inflammatory infiltrate in osteomyelitis can confound interpretation of biopsy specimens in the absence of immunohistochemistry. The FTIR data show that despite this histological mimicry, the underlying biochemical composition of the two tissues differs in measurable ways, particularly in phospholipid metabolism, protein secondary structure, and nucleic acid content. These differences reflect the distinct cellular mechanisms (rapid tumor proliferation vs. bacterial-driven inflammation) and exist at a molecular level that light microscopy does not resolve.
Integration with the diagnostic pathway: The authors propose a model where FTIR spectroscopy serves as a first-pass screening layer before or concurrent with histopathology. For centers with limited pathology expertise, this could provide an objective machine-readable signal. For referral centers, it could flag borderline cases for expedited molecular workup. The approach also aligns with the broader trend in oncology toward multimodal diagnostic integration, where no single test is definitive but converging evidence from multiple methods raises diagnostic confidence and speeds clinical decision-making.
The most substantial limitation of this study is its small sample size. With 27 ES patients and only 10 osteomyelitis patients, the cohort is underpowered for robust machine learning training and validation. The imbalance between classes (more than 2:1 ES to osteomyelitis ratio) is also a concern, as classifiers trained on imbalanced datasets can develop a systematic bias toward the majority class. While the authors used an external test set for validation rather than simple cross-validation, the absolute numbers in that test set are too small to draw definitive conclusions about real-world diagnostic accuracy.
Retrospective single-center design: The study is retrospective, meaning samples were not prospectively collected according to a pre-specified protocol, introducing potential selection bias. All samples came from a single institution, limiting geographic and demographic generalizability. Bone FTIR spectra can vary with patient age, sex, baseline bone mineral density, and prior treatment exposure (including antibiotics for osteomyelitis and prior chemotherapy for ES), none of which were systematically controlled or analyzed as covariates in this study.
Classification scope: The three-class model evaluated here (normal bone, ES, osteomyelitis) does not reflect the full differential diagnostic space a clinician faces. Bone lesions can also represent other primary bone sarcomas (osteosarcoma, chondrosarcoma), metastatic disease, or other infections. A clinically useful FTIR classifier would need to perform robustly across this broader disease spectrum, which requires substantially larger and more diverse training data.
Path to clinical translation: Future work should prioritize prospective multi-center cohort studies with pre-specified sample sizes informed by power calculations, inclusion of the full bone lesion differential, standardized FTIR acquisition protocols across multiple instrument platforms, and independent external validation on data collected at different institutions. Automating the spectral preprocessing and classification pipeline into a software tool that outputs a classification probability alongside confidence intervals would be a concrete step toward clinical usability. Pairing FTIR analysis with existing imaging radiomics scores could further improve discrimination without requiring additional tissue.