Cystoscopy is invasive and imperfect. The current gold standard for bladder cancer diagnosis requires inserting a camera into the bladder, which is uncomfortable, expensive, and can miss low-grade tumors (sensitivity only 61%). With a recurrence rate of 50 to 80%, patients must undergo this procedure every three to six months, making bladder cancer the most expensive cancer to treat per patient.
Voided urine cytology (VUC), the main noninvasive alternative, examines cells from urine under an optical microscope. While specificity is high (over 90%), sensitivity is only 20 to 80% overall and as low as 20 to 25% for low-grade tumors -- the most common type. Worse, results vary significantly between observers.
Biochemical biomarker tests such as NMP22, BTA, and FISH have been studied but remain insufficiently accurate or repeatable for widespread clinical adoption and are not currently recommended for routine diagnosis.
This study addresses an unmet need for a noninvasive, objective, and highly accurate test by combining atomic force microscopy (AFM) at nanometer resolution with machine-learning classification of cell surface features extracted from urine samples.
Cells from urine, imaged at nanoscale. Urine samples were collected from 43 individuals without bladder cancer and 25 patients with pathologically confirmed bladder cancer (14 low-grade, 11 high-grade) at Dartmouth-Hitchcock Medical Center and Cleveland Clinic. Cells were collected in a manner similar to standard VUC, then fixed, washed, freeze-dried on glass slides, and imaged with AFM.
Two subresonance tapping AFM modalities were used: PeakForce and Ringing mode. Ringing mode is 5 to 20 times faster than standard subresonance tapping and less prone to artifacts, making it more practical for clinical application. An optical microscope built into the AFM was used to locate cells randomly before imaging.
Two imaging channels were recorded for each cell: sample height and adhesion between the AFM probe and the cell surface. These two channels were selected because they are quantitatively robust across different AFMs, probes, temperature, and humidity conditions, making results reproducible across laboratories.
On average six cells per patient were imaged at 10 x 10 micron resolution. Eighteen out of 43 healthy controls and one low-grade cancer sample showed no cells in their urine. AFM revealed that many objects identified as cells by optical microscopy were actually non-cellular debris showing distinctive layered structures, demonstrating an inherent advantage of AFM imaging over conventional cytology.
Engineering surface descriptors applied to cells. Rather than feeding raw images directly into a classifier, the team extracted quantitative surface parameters traditionally used in materials engineering: roughness, directionality, fractal properties, and related measures. This reduced the dimensionality of the data substantially, avoiding the large training sets required by raw image methods.
Parameters were ranked by their segregation power using the Gini index from Random Forest models. Only the 8 to 10 parameters with the highest segregation power and low inter-parameter correlation were retained for classification. Adding more parameters actually decreased accuracy, because lower-ranked parameters contributed more noise than signal.
Three machine-learning classifiers were evaluated to confirm results were not algorithm-specific: Random Forest, Extremely Randomized Forest (both bootstrap unsupervised methods), and Gradient Boosting Trees (a supervised method). All three were chosen for their resistance to overfitting.
Training and testing were performed on independent cohorts, with all data from the same individual kept in one cohort to prevent information leakage between cells of the same patient. The dataset was split randomly 1,000 times at 50/60/70% training fractions, generating full statistical distributions of accuracy, sensitivity, specificity, and AUC across all possible splits.
Single-cell analysis: 80% accuracy. Analyzing just one cell per patient, the adhesion channel achieved approximately 80% accuracy with high statistical consistency across all training-testing splits and split percentages. The height channel produced almost no useful discriminative information, consistent with lower image resolution in that channel.
When the diagnosis was based on five cells (N=5, M=2 threshold -- cancer diagnosed if at least two of five cells test positive), accuracy rose to 94 plus or minus 1% across all three machine-learning methods and all 1,000 random splits. AUC reached 0.91 to 0.92.
Sensitivity and specificity trade-off. For the five-cell analysis, one operating point offered 81% sensitivity and 98% specificity, while another balanced both at approximately 91% sensitivity and 82% specificity. Both outperform currently available noninvasive methods such as VUC (sensitivity 20 to 80%) and biochemical markers like FISH (67 to 87% sensitivity, 72 to 96% specificity).
An anti-overfitting validation confirmed the results are genuine: when diagnoses were randomly shuffled and the same pipeline applied, accuracy fell to 53 plus or minus 10% (close to chance), confirming no classification artifacts exist in the method.
The cancer glycocalyx signature. The adhesion images capture nonspecific adhesion between the AFM probe and the cell surface glycocalyx -- the sugar-protein coating that envelops all cells. In malignant cells, the glycocalyx composition and spatial distribution changes in ways that alter the adhesion map at the nanoscale.
Unlike prior AFM studies that measured cell stiffness (elastic modulus) at a single point per cell, this method creates a full spatial map of adhesion across the entire cell surface. The most discriminative features describe the distribution of adhesion values across the surface, not just average adhesion, capturing subtle spatial patterning invisible to point-indentation methods.
Field carcinogenesis supports the multi-cell approach. The concept of field carcinogenesis holds that in bladder cancer, genetic and environmental risk factors affect the entire organ, not just the localized tumor. This means most cells collected from a cancer patient should carry some physical signature of cancer, explaining why randomly selected cells -- not just visually abnormal ones -- can serve as diagnostic material.
Compared to standard cystoscopy (AUC = 0.77) and VUC (sensitivity 20 to 80%), this method achieved statistically significant improvement (P less than 0.05), making it a potential first clinical application of AFM technology despite its 30-year history as a research tool.
Potential to eliminate unnecessary cystoscopies. Because the test uses urine collected non-invasively and analyzed objectively by machine learning, it could serve as a triage tool to identify which patients truly need cystoscopy. This would reduce patient discomfort, procedural complications, and the enormous cost associated with routine surveillance cystoscopy.
The method requires only a small number of randomly chosen cells (five in this study), a major departure from VUC which screens large numbers of cells looking for morphologically abnormal ones. This makes the approach faster and less dependent on finding rare malignant cells in the sample.
Beyond bladder cancer. The authors note that the same approach could apply to any cancer where cells can be collected from body fluids without invasive biopsy: upper urinary tract cancer, urethral cancer, colorectal and gastrointestinal cancers, cervical cancer, aerodigestive cancers, and others. It could also extend to monitoring cell responses to drugs (nanopharmacology).
A patent application was filed by Tufts University for the AFM-machine learning method. Larger patient cohorts are needed before clinical introduction, but the framework represents a potentially new direction for biomedical diagnostic imaging grounded in physical rather than biochemical properties of cells.
A statistically significant advance over cystoscopy. Tested on 68 patients (43 controls, 25 cancer), the method achieved 94% accuracy and AUC of 0.91 to 0.92 -- statistically significantly better than the invasive clinical standard, cystoscopy (AUC = 0.77, P less than 0.05), while being completely noninvasive.
Results were consistent across three independent machine-learning algorithms, 1,000 random train-test splits, and multiple training-set size fractions. The robustness of findings across all these variables suggests the signal detected is real and stable, not an artifact of a particular analytic choice.
The adhesion channel of AFM uniquely captures cell surface information unavailable from optical, spectroscopic, or biochemical methods. This makes the AFM-derived features complementary to existing diagnostic tools rather than simply duplicating what other methods already measure.
The study represents a proof-of-concept on a moderate-sized cohort. Expansion to larger multicenter cohorts will be needed to validate the 94% accuracy figure and confirm performance across the full spectrum of tumor grades and stages before clinical deployment.