For over a century, cancer diagnosis has relied on examining thin two-dimensional (2D) slices of tumor tissue under a microscope. A pathologist picks a few representative sections from a biopsy core or surgical specimen, and these guide all subsequent treatment decisions. But tumors are not flat -- they are inherently three-dimensional objects with complex, spatially varying architecture, and a single 2D slice captures only a tiny fraction of this information.
Research has shown that diagnoses become more accurate when multiple tissue levels are examined rather than a single slice. This is because prostate tumors and many other cancers exhibit significant tissue heterogeneity -- the same tumor block can contain regions of highly aggressive Gleason grade 4 cancer next to relatively indolent Gleason grade 3 tissue. A 2D slice that happens to miss the aggressive region will underestimate the tumor's true risk, potentially leading to undertreated disease and worse outcomes.
Three-dimensional (3D) tissue imaging technologies have been developed to capture the full volumetric structure of tumors. These include open-top light-sheet microscopy (OTLS), which uses fluorescent staining to create detailed 3D reconstructions of cleared tissue, and microcomputed tomography (microCT), which builds 3D volumes from X-ray projections. Both produce gigavoxel datasets with cellular-level resolution -- but analyzing these massive images manually is impractical even for the most experienced pathologists.
This study introduces MAMBA (Modality-Agnostic Multiple instance learning for volumetric Block Analysis) -- the first end-to-end deep learning platform specifically designed to analyze 3D pathology images and predict patient outcomes without requiring pathologists to manually annotate every structure in the volume.
MAMBA is a weakly supervised deep learning system, meaning it learns from patient-level labels only (such as whether a patient experienced cancer recurrence within five years) without requiring pathologists to annotate individual structures like glands, cells, or tumor boundaries. This is critical because manual annotation of 3D tissue volumes at cellular resolution would be prohibitively time-consuming and is the main barrier to deploying AI in 3D pathology.
The pipeline works in three steps. First, the raw 3D tissue volume is segmented to separate tissue from background, then divided into smaller 3D cuboids (patches) -- three-dimensional chunks rather than flat 2D slices. This subdivision makes computation tractable while preserving 3D spatial relationships. Second, each cuboid is processed through a pretrained deep learning feature encoder that converts the visual content of each patch into a compact numerical representation using transfer learning -- a technique where networks pretrained on large image databases are adapted to the medical domain.
The third step uses an attention-based aggregation module to intelligently combine all the cuboid-level features into a single patient-level prediction. The attention mechanism automatically learns to weight which 3D patches are most informative for predicting outcomes, effectively identifying the most diagnostically important regions of the tissue without any human guidance. This produces a final risk score predicting whether the patient is likely to experience biochemical recurrence.
A key design feature is that MAMBA is modality-agnostic: it works with any 3D imaging technology, whether OTLS, microCT, or future modalities, and can process the same volume as either 2D slices or 3D cuboids, enabling direct performance comparisons between approaches.
The platform was tested on two independent prostate cancer cohorts from different hospitals. The University of Washington (UW) cohort contained 50 simulated core needle biopsies from prostatectomy specimens, imaged with OTLS at 1 micrometer per voxel resolution -- sufficient to see individual cell nuclei and gland architecture clearly. The Brigham and Women's Hospital (BWH) cohort contained 45 prostatectomy specimens imaged with microCT at 4 micrometers per voxel.
Both cohorts used five-year biochemical recurrence (BCR) as the clinical endpoint: patients were classified as high-risk if their PSA rose above threshold within five years of prostatectomy and low-risk otherwise. Each cohort was exactly balanced with equal numbers of high-risk and low-risk patients, eliminating any potential class imbalance bias. Performance was assessed using five-fold cross-validation, repeated five times with different data splits.
Before testing on real tissue, MAMBA was validated on simulated 3D phantom datasets -- synthetic volumes with mathematically defined cell shapes representing two different tissue types. This allowed systematic testing of whether MAMBA could reliably detect morphological differences in controlled settings where the ground truth was known exactly, serving as proof-of-concept before applying the method to complex real biological data.
The comparison framework was thorough: MAMBA was tested using data from a single 2D plane (emulating standard 2D pathology), all 2D planes from the full volume, and 3D cuboids from the full volume. This design allowed direct quantification of how much prognostic information is gained by moving from 2D to 3D analysis.
In the OTLS cohort, MAMBA using 3D cuboids from the whole volume achieved an AUC of 0.857, significantly outperforming analysis of all 2D planes from the same volume (AUC 0.829, p less than 0.05) and far exceeding the single-slice approach (AUC 0.792, p less than 0.01). The improvement was also clearly reflected in Kaplan-Meier survival curves, where 3D MAMBA created much better separation between predicted high-risk and low-risk patient groups than single-plane analysis.
In the microCT cohort, the advantage of 3D over 2D processing was even more pronounced. Whole-volume 3D cuboids achieved an AUC of 0.739, compared to 0.631 for 2D planes and 0.567 for a single plane -- a difference of more than 0.17 AUC points between 3D and single-slice approaches. The gap between 3D and 2D was also statistically larger for microCT than OTLS, suggesting 3D processing is especially valuable when image resolution is lower.
Ablation analysis confirmed a consistent upward trend: the more of the tissue volume incorporated into the analysis, the better the prognostic performance -- regardless of whether data was treated as 2D planes or 3D cuboids. This directly demonstrates that prostate cancer prognosis is better predicted when more of the tumor's spatial heterogeneity is captured, supporting the biological rationale for 3D pathology.
Cross-modal experiments -- training on one modality and testing on the other -- achieved AUCs of 0.676-0.725, suggesting MAMBA learns genuine morphological patterns that generalize across imaging technologies rather than overfitting to scanner-specific artifacts. This is an important finding for clinical translation, where different institutions will use different 3D imaging systems.
To understand what MAMBA was learning, the researchers used integrated gradient (IG) analysis -- a technique that assigns each 3D patch a score reflecting how much it influenced the model's risk prediction, with positive scores (red) indicating regions that push the prediction toward high risk and negative scores (blue) indicating regions associated with low risk.
Analysis of the highest-scoring patches revealed a biologically meaningful pattern: patches from the high-risk group showed infiltrative carcinoma resembling Gleason grade 4, including cribriform architecture (a pattern where glands fuse together abnormally). Patches from the low-risk group showed predominantly large, benign-appearing glands. Middle-scoring patches contained mixtures of Gleason grades 3 and 4. These findings align directly with what pathologists would predict based on known prostate cancer biology.
The correlation between each patient's average IG score and their MAMBA-predicted risk was extremely strong (Pearson r = 0.94, p less than 0.0001 in OTLS; r = 0.95 in microCT), confirming that the model is consistent in its attribution of risk. Higher-risk patients had proportionally more high-scoring patches throughout their tissue volume, while lower-risk patients had more low-scoring patches -- exactly the expected relationship.
Critically, the analysis showed that partial-volume analysis can miss important prognostic regions. When only 15% of a tissue volume was randomly sampled (emulating biopsy subsampling), AUC spread was substantial (median 0.806, with a range spanning 0.112) in 50 repeated trials, and 80% of random samples performed worse than whole-volume analysis. This quantifies the sampling bias risk of standard 2D pathology and reinforces the case for 3D analysis.
A common concern about 3D analysis is that it requires far more computational resources than 2D approaches. MAMBA addresses this directly: while 3D patches are computationally heavier to encode individually, the total number of 3D patches is roughly 30 times fewer than the equivalent number of 2D patches for the same tissue volume. This is because each 3D cuboid encompasses the depth dimension that would otherwise require dozens of separate 2D slices.
The net result is that network training and prediction time for 3D processing is up to 40 times faster than 2D alternatives (3D train: 0.025 seconds per epoch per sample vs. 2D: 0.98 seconds). Memory requirements for stored features are similarly reduced by approximately 30-fold. This means 3D analysis is not only more accurate but also more computationally tractable than processing the equivalent data as stacked 2D images.
As 3D imaging technologies continue to improve -- with higher resolutions, larger fields of view, and lower acquisition costs -- the datasets they generate will grow substantially. The computational efficiency advantage of MAMBA's 3D approach will become even more important as dataset sizes increase, making it well-positioned to scale with the field.
The most direct clinical application is improving prostate cancer risk stratification after surgery. Current pathological grading relies on a pathologist examining a few 2D slides from the specimen -- inevitably missing heterogeneous regions. MAMBA's ability to analyze the full tissue volume could provide a more complete picture of disease aggressiveness, potentially identifying patients at higher recurrence risk who would benefit from early adjuvant therapy.
Beyond prostate cancer, MAMBA was designed as a general platform. Its modality-agnostic architecture means the same computational framework can be applied to 3D tissue images from pancreatic cancer, colorectal cancer, melanoma, and kidney cancer -- all areas where 3D imaging is beginning to reveal heterogeneity invisible to 2D sections. As 3D imaging costs decrease and protocols become standardized, MAMBA could become a foundational tool for clinical computational pathology.
For pathologists, MAMBA functions as a decision support tool rather than a replacement. The interpretability heatmaps highlight specific 3D regions driving the risk prediction, allowing pathologists to visually confirm and contextualize the AI's assessment. This transparency is critical for clinical trust -- physicians need to understand why the AI is flagging a region as high-risk, not just that it is.
The platform also opens the door to discovering genuinely new 3D morphological biomarkers -- features of tissue architecture that only become apparent when examined in three dimensions and have no equivalent in 2D pathology. These might include the three-dimensional shape of gland openings, the volumetric distribution of tumor versus benign tissue, or spatial patterns of immune cell infiltration that 2D slices sample only partially.
MAMBA represents the first end-to-end deep learning platform for 3D pathology that achieves superior patient prognostication compared to 2D alternatives while remaining computationally efficient and interpretable. Across two independent prostate cancer cohorts imaged with different modalities, 3D whole-volume analysis consistently outperformed single-slice and multi-plane 2D approaches for predicting five-year biochemical recurrence after radical prostatectomy.
The consistent finding that more tissue volume equals better prediction -- combined with the quantified risk of sampling bias in partial-volume analysis -- provides a strong evidence-based argument for transitioning prostate cancer pathology toward 3D imaging and analysis. The interpretability analyses confirm that MAMBA's predictions are grounded in established morphological biomarkers like Gleason grade patterns, supporting biological validity.
The main limitation is the small cohort size (45-50 patients per cohort), reflecting the high cost of 3D tissue imaging and the scarcity of long-term outcome data for 3D-imaged specimens. Despite representing an unprecedented dataset size for 3D pathology, these numbers are too small to identify novel 3D biomarkers or perform robust subgroup analyses. Larger cohorts are needed to fully validate and expand on these findings.
Future work should focus on fine-tuning the feature encoders specifically for 3D tissue images rather than relying on transfer learning from natural image databases, optimizing aggregation architectures to exploit spatial context within volumes, and extending validation to prospective patient cohorts. With continued development, MAMBA has the potential to bring the diagnostic and prognostic power of 3D tissue morphology into routine clinical practice.