Prostate cancer is the second most commonly diagnosed cancer in men worldwide. Current standard screening tools -- including PSA blood tests, digital rectal exams, and transrectal ultrasound -- have significant limitations in accuracy, resolution, and tissue chemistry information.
Photoacoustic (PA) spectroscopy is a hybrid imaging technique that fires short pulses of near-infrared laser light into tissue. Biological molecules absorb this light, convert it to heat, and expand rapidly -- generating ultrasound waves that are then recorded by sensors. The pattern of these acoustic signals reveals the chemical composition of the tissue.
Because different biological molecules absorb light at different wavelengths, PA spectroscopy can simultaneously map the distribution of haemoglobin, lipids, and collagen inside tissue -- information that standard ultrasound and MRI cannot provide. This physio-chemical data may reveal cancer-related changes in tissue microstructure.
The challenge is that PA spectroscopy generates enormous amounts of complex signal data. Machine learning (ML) offers a systematic way to analyze these signals and extract the patterns that distinguish cancerous tissue from healthy tissue, making automated diagnosis feasible.
The study used 22 human prostate tissue samples from patients who had undergone radical prostatectomy (surgical removal of the prostate) without prior therapy. Samples were transported on ice and tested within two hours of removal to preserve tissue integrity.
PA measurements were taken at 97 sites across the tissue samples using needle probes inserted at locations guided by expert pathologists. Follow-up pathological examination confirmed that 47 sites were cancerous and 50 were healthy, providing the ground-truth labels for training the machine learning models.
Laser pulses swept across wavelengths from 690 to 950 nm and 1200 to 1690 nm -- covering the key absorption regions of deoxyhaemoglobin, oxyhaemoglobin, lipids, collagen, and water. Signals were averaged 128 times to improve the signal-to-noise ratio.
The raw PA signals were preprocessed using the Pwelch method (a standard power spectrum estimation technique) to convert time-domain signals into frequency-domain power spectra, creating a two-dimensional data matrix of 77 wavelengths by 191 frequency points for each measurement site.
To make sense of the 77-wavelength data, the researchers used hierarchical clustering via the UPGMA (Unweighted Pair Group Method with Arithmetic Mean) algorithm. This groups wavelengths with similar power spectrum patterns -- effectively identifying which parts of the spectrum correspond to which biological molecules.
The clustering revealed six distinct wavelength groups corresponding to known biological chromophores: deoxyhaemoglobin (680-790 nm), oxyhaemoglobin (800-940 nm), lipid type 1 (1200-1250 nm), collagen (1260-1390 nm), water (1400-1640 nm), and a second lipid group (1650-1680 nm).
The researchers also calculated Pearson correlation coefficients (PCCs) between all pairs of wavelengths for each tissue sample, creating a 77x77 correlation matrix. These were visualized as network maps where nodes represent wavelengths and connections represent strong correlations.
Comparing correlation networks between healthy and cancerous tissues revealed a key finding: cancerous tissue shows far more inter-group correlations -- more edges, larger nodes, and more labeled wavelength clusters -- suggesting that cancer disrupts the normal separation between different molecular distributions in tissue.
The PA physio-chemical spectra showed clear differences between healthy and cancerous prostate tissue. In the 680-940 nm range (dominated by haemoglobin), cancerous tissue produced significantly stronger signals, consistent with the well-known phenomenon of angiogenesis -- the growth of new blood vessels that tumours require to sustain themselves.
In the 1200-1370 nm range, where lipids and collagen are the primary signal sources, cancerous tissue also showed greatly amplified signals. This indicates that collagen and lipid concentrations rise as prostate cancer develops, reflecting changes in the structural scaffold and cell membrane composition of tumour tissue.
The correlation analysis showed that the intergroup association between haemoglobin-related wavelengths (W1-W2) and lipid/collagen wavelengths (W3-W6) was significantly higher in malignant tissue. In normal tissue, these molecular systems are relatively independent; in cancer, they become more uniformly and chaotically distributed.
This chemical reorganization makes intuitive biological sense: as tumour metabolism becomes irregular and the microenvironment changes, the normally distinct spatial distribution of collagen fibres and lipid-rich cell membranes breaks down, leading to increased spectral correlation. These changes are directly detectable by PA spectroscopy.
Four machine learning classifiers were applied to the extracted PA spectral features: Support Vector Machine (SVM), Naive Bayes (NB), Decision Tree C4.5, and Linear Discriminant Analysis (LDA). All four were trained and evaluated using 10-fold cross-validation repeated three times.
C4.5 achieved the highest accuracy at 97.3%, followed by SVM at 96.8%, LDA at 95.8%, and Naive Bayes at 95.2%. All four methods exceeded 92% on both sensitivity and specificity, demonstrating consistently strong performance across the full dataset of 97 tissue sites.
For comparison, standard clinical imaging tools showed much lower accuracy on the same diagnostic task: multi-parametric ultrasound achieved roughly 71.7% accuracy, while MRI achieved approximately 80%. The PA-based machine learning approach outperformed both by substantial margins.
The LDA model worked by identifying specific frequency points in the power spectrum that maximally separated healthy from cancerous tissue. By combining the mapping values at multiple well-chosen frequency points, the model achieved its high accuracy even with only 97 data points -- a very small dataset by machine learning standards.
Linear Discriminant Analysis (LDA) works by finding the linear combination of input features that maximally separates the two classes (cancer vs. healthy). It reduces the high-dimensional spectral data to a single mapping value -- positive for cancer, negative for healthy -- using the 77 wavelength power values at each selected frequency point.
Support Vector Machines (SVM) find the optimal boundary (hyperplane) between classes in high-dimensional feature space, maximizing the margin between the two groups. SVMs are particularly well-suited to problems where the number of features is large relative to the number of samples.
The C4.5 Decision Tree builds a hierarchical series of yes/no questions about feature values, using information gain ratio to select the most discriminating features at each split. It is interpretable -- the resulting tree explicitly shows which wavelength features most distinguish cancer from healthy tissue.
Naive Bayes calculates the probability of cancer given each feature independently, then combines these probabilities assuming statistical independence between features. Despite this simplifying assumption, it performed competitively, suggesting that the extracted features are relatively informative even when considered individually.
The study demonstrates that photoacoustic spectroscopy combined with machine learning can detect prostate cancer with accuracy comparable to biopsy-level assessment, while being far less invasive. The PA measurements were taken with needle probes, but the researchers note the potential for fully noninvasive external imaging configurations.
A key clinical advantage of PA spectroscopy over conventional imaging is its ability to provide simultaneous chemical and structural information. Rather than relying solely on visual appearance, it directly measures the concentrations of cancer-relevant molecules like haemoglobin (a marker of angiogenesis) and collagen.
The dataset in this study was intentionally small (97 measurements from 22 prostates) but was validated through rigorous 10-fold cross-validation. The authors acknowledge that larger datasets would further refine the models and better characterize the statistical properties of the spectral features.
Future work should focus on distinguishing benign from malignant tumours (this study only classified cancer vs. normal tissue), developing real-time signal processing pipelines, and ultimately testing the approach on in vivo tissue rather than ex vivo surgical specimens.
This study establishes a proof of concept that photoacoustic physio-chemical spectroscopy can accurately classify prostate tissue as cancerous or normal when analyzed with machine learning. The approach is both physically grounded -- in the known optical absorption properties of biological molecules -- and practically effective.
All four tested ML classifiers (SVM, Naive Bayes, C4.5, LDA) exceeded 94% accuracy on the 97-site dataset, with C4.5 reaching 97.3%. These results substantially surpass standard clinical imaging accuracy, validating the diagnostic power of the physio-chemical information captured by PA spectroscopy.
The finding that cancer increases correlations between haemoglobin, collagen, and lipid signals provides a biologically interpretable signature of malignancy. This aligns with known cancer biology -- tumour angiogenesis, altered extracellular matrix remodeling, and disrupted lipid metabolism all contribute to these spectral changes.
The method's combination of high accuracy, chemical specificity, and potential for minimal invasiveness positions PA spectroscopy as a promising technology for next-generation prostate cancer screening and diagnosis, pending further validation in larger clinical studies.