For early-stage, hormone receptor-positive breast cancer patients with no lymph node involvement, predicting who will experience recurrence remains difficult. Standard molecular tests like Oncotype DX (which measures 21 genes) and MammaPrint (a 70-gene panel) are widely used, but their predictive power is modest, with reported AUCs of only 0.69 and 0.59 respectively.
A newer test, OncoMasTR, uses just three master transcription regulators combined with tumor size and lymph node status to stratify patients as low or high risk. While analytically validated, it and similar tests still share a fundamental limitation: they analyze bulk tumor tissue and lose the spatial arrangement of cells and their surrounding microenvironment.
Discordance between different molecular tests -- where the same patient receives conflicting risk classifications from different assays -- underscores the unsettled landscape of recurrence prediction and suggests no single test captures the full biological picture.
This study explored whether Fourier-transform infrared (FTIR) chemical imaging of breast tissue, when analyzed with deep learning, could offer a complementary label-free approach to predicting recurrence while preserving both the spatial and biochemical context of the tumor microenvironment.
Fourier-transform infrared (FTIR) spectroscopy measures how molecules in tissue absorb infrared light across a range of wavelengths. Different molecular bonds -- including those in proteins, lipids, nucleic acids, and carbohydrates -- have characteristic spectral signatures. When combined with a microscope, FTIR produces spatially resolved images where each pixel contains a full biochemical spectrum.
Unlike standard histology (which uses dyes like hematoxylin and eosin to highlight tissue structure), FTIR is label-free and non-destructive. It does not require staining, which preserves the tissue for other uses and removes dye-related variability. The result is a detailed chemical map of the tissue without any added reagents.
FTIR captures signals from molecular modifications that reflect disease biology at the biochemical level, including DNA methylation, phosphorylation, and glycosylation. These changes can precede morphological alterations visible in standard pathology, potentially enabling earlier detection of prognostic signals.
The key challenge for FTIR imaging is data dimensionality: each image contains hundreds of spectral channels measured spatially, making it a high-dimensional dataset that standard image analysis methods struggle to handle efficiently. This study applied deep learning specifically designed for this type of hyperspectral data.
The study used formalin-fixed, paraffin-preserved tissue microarray (TMA) specimens from 144 breast cancer patients diagnosed at Malmo University Hospital in Sweden. After excluding samples with core loss, 139 patients contributed usable imaging data, of whom 29 experienced disease recurrence during follow-up.
The patient population was diverse, with 125 of 144 patients being ER-positive and 51 being HER2-positive. The mean age at diagnosis was 65 years, and patients were treated with mastectomy or wide local excision without neoadjuvant therapy, ensuring that tissue samples reflected pre-treatment tumor biology.
FTIR imaging was conducted using a focal plane array detector at the University of Exeter, imaging each TMA core over the fingerprint region (1000-1800 cm-1), which contains the most diagnostically informative molecular absorption bands. Serial sections from the same blocks were stained with hematoxylin and eosin to provide matched reference images.
Spectral preprocessing removed background signals including resonant Mie scattering artifacts and paraffin wax contributions. Pixels not containing tissue were identified and zeroed out using k-means clustering, leaving only tissue spectra for analysis.
The study tested four distinct analytical strategies. The first used traditional partial least squares discriminant analysis (PLSDA) on spectra extracted from image segments, representing the conventional approach in chemical imaging. The second used transfer learning from pretrained 2D CNNs (VGG16, Xception, InceptionV3, DenseNet121) applied to compressed 3-channel representations of the spectral images.
The third approach trained standard 2D CNN architectures -- including AlexNet, LeNet5, VGG16, and Xception -- directly on the full hyperspectral images from scratch, without compressing the spectral dimension. This tested whether coupling spectral and spatial information within a single learning pass would improve performance.
The fourth and most novel approach replaced 2D convolutions with depth-wise separable convolutions (DSCNNs) that separately process spatial and spectral dimensions within each layer. Two variants were designed: Type 1 DSCNNs process spatial information first and then mine spectral features, while Type 2 DSCNNs exploit both dimensions simultaneously for deeper information extraction.
For all deep learning models, the dataset was split 60/20/20 for training, validation, and testing, with five-fold cross-validation used to assess model robustness. Class weighting addressed the imbalance in the dataset, where only 21% of patients experienced recurrence.
Traditional PLSDA models applied to segmented chemical image clusters performed poorly, failing to identify any spectral cluster that reliably predicted recurrence. Transfer learning models using pretrained 2D CNNs with compressed spectral representations (VGG16, Xception, DenseNet121) similarly failed, achieving AUCs near 0.50 -- equivalent to random guessing.
When 2D CNN architectures were trained directly on the full hyperspectral images, performance improved substantially. The champion model, a modified LeNet5 architecture, achieved a ROC AUC of 0.639 plus or minus 0.081 for breast cancer recurrence prediction. The much deeper InceptionV3 came second at 0.577.
The best depth-wise separable CNN model (Model 2, Type 1 DSCNN) achieved an AUC of 0.628 plus or minus 0.095, closely matching the LeNet5 result. Notably, shallower and simpler models consistently outperformed deeper, more complex architectures -- likely because the small dataset size (139 patients) limits the benefit of additional model capacity.
When the prediction task was redefined as classifying survival within 7 years (rather than any recurrence), model performance improved to AUC of 0.68, suggesting that refining the clinical endpoint definition -- for example excluding deaths from non-cancer causes -- may help the model learn more coherent patterns.
Standard H&E-based digital pathology assesses tissue morphology -- the visual shape, size, and arrangement of cells. FTIR chemical imaging adds a fundamentally different layer: the intrinsic biochemistry of the tissue, including protein folding states, nucleic acid modifications, and lipid composition, captured with spatial resolution.
Because biochemical changes such as altered methylation, phosphorylation, and metabolic reprogramming occur before they produce visible morphological changes, spectral data may encode prognostic information that is simply invisible to traditional histological analysis. This is supported by the observation that this study's models slightly outperformed equivalent H&E-based histomorphometric approaches at the same predictive task.
A critical finding is that simply compressing spectral information into a few channels (as done in the transfer learning approach) destroyed predictive value. Successful models needed to process both spatial and spectral information together, confirming that recurrence-relevant signals reside in the combined spatial-chemical landscape of the tissue.
The finding that PLSDA on individual spectral clusters failed entirely also reinforces the view that tissue biochemistry predicting recurrence is not localized to a single cell type or region. Rather, it appears to reflect emergent properties of the tissue microenvironment that require whole-image analysis to capture.
This study presents proof-of-concept evidence that chemical imaging combined with deep learning can predict breast cancer recurrence with performance comparable to currently approved genomic tests. This positions FTIR-based analysis as a potential label-free, reagent-free alternative to molecular assays.
Practical advantages of this approach include the use of standard paraffin-preserved tissue already collected during surgery, no need for specialized reagents or PCR analysis, and potential integration into routine pathology workflows. The technology could serve as a first-stage triage tool, flagging patients who warrant full genomic testing.
The small cohort size (139 patients) is the central limitation of this study, directly constraining model performance and increasing uncertainty in AUC estimates. Larger, multicenter validation studies with thousands of patients would be required to establish whether these models generalize across patient populations, clinical settings, and instrumentation platforms.
Future work could extend the approach to predict overall survival probability, response to specific therapies, and endpoints in other cancer types. The authors also note the potential for intraoperative applications, similar to stimulated Raman imaging systems that are already being tested for real-time surgical decision support.