Diffuse Large B-Cell Lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma, accounting for over a third of all NHL cases and affecting more than 20,000 patients annually in the United States. Despite its prevalence, DLBCL presents a fundamental clinical challenge: treatment outcomes following the standard R-CHOP regimen (rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone) are highly variable, and approximately 30% of patients are either refractory or relapse after initial therapy. The disease is fatal without treatment, yet identifying which patients are at highest risk remains difficult.
Current prognostic tools: The most widely used prognostic tool is the National Comprehensive Cancer Network International Prognostic Index (NCCN-IPI), which incorporates five clinical variables: age, lactate dehydrogenase (LDH), extra-nodal sites of involvement, Ann Arbor stage, and ECOG performance status. While this index stratifies patients into good, intermediate, and poor-risk categories, it is insufficient to guide individualized therapeutic decision-making. Gene expression profiling (GEP) adds information by classifying DLBCL into cell-of-origin (COO) subtypes - germinal center B-cell (GCB), activated B-cell (ABC), or indeterminate - based on the phase of B-cell development the tumor resembles. The GCB subtype is associated with more favorable outcomes than non-GCB, and the Hans algorithm using immunohistochemistry is the most widely applied COO classification method in clinical practice, though it shows imperfect concordance with molecular GEP.
The morphologic gap: Despite DLBCL displaying varying morphologies under the microscope - ranging from centroblastic (multiple peripheral nucleoli, vesicular chromatin) to immunoblastic (abundant cytoplasm, single prominent nucleolus) and anaplastic variants - no morphologic features have been consistently demonstrated to correlate with prognosis. Multiple prior studies failed to show that manual morphologic classification reliably predicts outcomes. This gap motivates the hypothesis that automated computational analysis could identify quantitative shape and size features of tumor nuclei that have eluded human pathologists.
This paper introduces DLBCL-Morph, a publicly available annotated dataset from Stanford Hospital designed to enable exactly this kind of computational morphologic analysis, paired with a demonstration that deep learning-derived nuclear geometry features do carry measurable prognostic signal.
The DLBCL-Morph dataset was constructed from a cohort of 209 patients with de novo, CD20-positive DLBCL treated with curative intent using R-CHOP or R-CHOP-like immunochemotherapy at the Stanford Cancer Institute. These patients were part of a prior study with defined clinicopathologic inclusion criteria, and all clinical data was de-identified in compliance with an IRB-approved protocol that waived the requirement for individual patient consent since the work had no impact on clinical care.
Tissue microarray construction: Formalin-fixed, paraffin-embedded (FFPE) tumor samples from each patient were assembled into tissue microarrays (TMAs) using 0.6-mm diameter core samples in duplicate, producing 0.4-micron thick FFPE sections. Duplicate cores per patient provide built-in replication for quality assessment. Seven distinct TMAs were produced, each stained with six different preparations: hematoxylin and eosin (H&E), which reveals cell morphology, plus immunohistochemical stains for five oncoproteins - CD10, BCL6, MUM1, BCL2, and MYC. These five markers are clinically significant because overexpression of MYC and BCL2 proteins (double-expressor phenotype) and concurrent chromosomal translocations in MYC/BCL2/BCL6 genes (double-hit lymphoma) are both associated with aggressive clinical course and poor outcomes under R-CHOP, and are now standard-of-care molecular assessments per WHO guidelines.
Digital scanning at 40x: All 42 TMA slides (7 TMAs x 6 stains) were scanned at 40x magnification (0.25 micrometers per pixel) on an Aperio AT2 scanner (Leica Biosystems) in ScanScope Virtual Slide (SVS) format. This high magnification captures fine cellular detail necessary for nuclear morphology analysis. Each SVS file contains the full TMA scan, a slide label image identifying the TMA number and stain, a macro camera image of the whole slide, and a thumbnail.
Clinical data: For each of the 209 patients, the dataset provides overall survival (OS, measured from end of treatment to death or last follow-up, in years), follow-up status (FUS, binary indicator of whether the patient was deceased at last follow-up), and a broad range of clinical and molecular variables including age and MYC/BCL2/BCL6 gene translocation status. All categorical variables in the cleaned CSV are converted to numerical values to facilitate machine learning applications.
A critical methodological challenge in working with TMA data is that not all tissue within a core is representative of DLBCL. Some cores are partially or entirely missing from the TMA grid. Others contain areas of tissue with very few or no tumor cells - stromal regions, necrosis, or processing artifacts - that would confound downstream analysis if included. To address this, expert pathologists manually annotated rectangular regions-of-interest (ROIs) on every core across all TMAs and all 6 stains, at 40x magnification. Each annotation identifies the portion of the core that accurately represents DLBCL tumor tissue.
Annotation storage format: The pixel coordinates of each ROI rectangle (upper-left and lower-right corners at 40x magnification) are stored in annotations.csv, linked to patient_id, TMA_id, and stain. Cores that are entirely missing or unrepresentative received no ROI annotations and are excluded from analysis. This explicit annotation of representative tissue is a significant quality control step that distinguishes DLBCL-Morph from datasets relying solely on core-level metadata.
Patch extraction: From within each annotated ROI, non-overlapping 224x224-pixel patches were extracted uniformly at 40x magnification, starting from the top-left corner of the rectangle and proceeding to the bottom-right. Patches consisting mostly of white background with little tissue content were omitted. The 224x224 pixel size is standard for deep learning image classification models (matching the input size of widely used architectures such as ResNet and VGG), making these patches directly usable as model inputs without additional preprocessing. Patches are stored in PNG format, organized by stain and patient_id. In total, 195 out of 209 patients have at least one patch from at least one stained TMA, while some patients lack patches for specific stains when cores were missing or not covered by ROIs.
This layered quality control - expert ROI annotation followed by intelligent patch filtering - ensures that downstream nucleus segmentation and feature computation operate exclusively on tissue areas that pathologists have confirmed as representative of DLBCL, substantially reducing noise in the morphologic measurements.
The centerpiece of the computational pipeline is tumor cell nucleus segmentation using HoVer-Net, a deep learning model developed for simultaneous segmentation and classification of nuclei in multi-tissue histology images. HoVer-Net was applied independently to each 224x224 patch extracted from H&E stained TMAs to identify every individual cell nucleus in the patch and assign it a cell type label.
HoVer-Net architecture: HoVer-Net uses a neural network based on a pretrained ResNet-50 backbone to extract image features, which are then processed through three parallel branches. The nuclear pixel (NP) branch determines whether each pixel belongs to a nucleus or to background. The HoVer branch predicts, for each nucleus pixel, the horizontal and vertical distances to the centroid of its parent nucleus - this clever approach uses spatial geometry to separate touching or overlapping nuclei, which would otherwise be merged into a single incorrectly segmented object. The nuclear classification (NC) branch classifies each nucleus pixel into one of five cell type categories: neoplastic, non-neoplastic epithelial, inflammatory, connective, or dead. Aggregating NC branch outputs across all pixels in a segmented nucleus assigns a single class label per nucleus.
Neoplastic nucleus isolation: For the purpose of morphologic feature computation, only nuclei classified as neoplastic (tumor cells) are retained. Each neoplastic nucleus is saved as a separate binary image - a two-dimensional array where pixel values indicate whether each position belongs to the nucleus or background. These binary segmentation images capture the precise shape and size of every individual tumor nucleus and are stored in Numpy NPY format. They are organized in the Cells folder by patient_id and patch_id, with cell numbers that are non-consecutive since all non-tumor cells are discarded per patch.
The HoVer-Net approach is particularly well-suited to lymphoma tissue, where densely packed tumor nuclei frequently touch one another. The horizontal/vertical distance prediction strategy for separating touching nuclei addresses one of the primary failure modes of simpler segmentation approaches such as watershed-based methods, which can struggle when cell boundaries are ambiguous in tightly packed tissue.
For each segmented neoplastic nucleus, the pipeline computes a set of geometric features from the binary segmentation image. These hand-crafted features encode the size and shape of the nucleus in a structured, interpretable form, as an alternative to relying solely on raw pixel data passed into end-to-end deep learning models. The authors argue that while end-to-end models may not require such features, prognostic models using geometric descriptors produce more explainable results and more clearly indicate which morphologic properties are prognostically informative.
Rectangle and ellipse fitting: A minimum-area rotated rectangle is fit to each binary nucleus mask. The dataset records the rectangle center coordinates (rectCenter), dimensions as width and height (rectDimension), and rotation angle in degrees (rotate_angle, ranging from -90 to 0). A rotated ellipse is also fit to each nucleus, providing the ellipse centroid (ellip_centroid), semi-axis lengths (shortAxis, longAxis), perimeter (ellip_perimt), and area (ellip_area). The rotation angle convention resets at -90 degrees for axis-aligned rectangles and approaches 0 as the rectangle rotates clockwise.
Feret diameters and convex hull: The maximum and minimum Feret diameters are computed for each nucleus. The Feret diameter in a given direction is the distance between two parallel tangent lines perpendicular to that direction. The maximum Feret diameter (maxDiameter) and the angle at which it is achieved (maxAngle, between -180 and 180 degrees) quantify the longest dimension of the nucleus, while minDiameter and minAngle capture the narrowest cross-section. The convex hull area (hull_area) - the area of the smallest convex polygon enclosing the nucleus - captures irregular boundary protrusions.
Derived shape ratios: Six derived features summarize nucleus geometry as dimensionless ratios. The elongation shape factor (esf) equals shortAxis/longAxis, equaling 1 for perfectly circular nuclei and decreasing toward 0 for elongated cells. The circularity shape factor (csf) equals 4*pi*ellip_area/ellip_perimt^2, also equaling 1 for circles and decreasing for irregular or elongated shapes. sf1 equals shortAxis/maxDiameter, sf2 equals minDiameter/maxDiameter, and elongation equals maxDiameter/minDiameter - all ratios measuring relative elongation. Convexity equals ellip_area/hull_area, measuring the compactness of the nucleus outline relative to its convex hull. All geometric features are stored per-nucleus in cell_shapes.csv, linked to patient_id, patch_id, and cell number.
To validate that the geometric features in DLBCL-Morph contain clinically useful information, the authors performed survival regression on the 170 patients for whom H&E stained TMA patches were available (out of 209 total). For each geometric feature computed per nucleus, the mean and standard deviation across all nuclei belonging to each patient were calculated, producing patient-level summary statistics. These were then combined with clinical variables for survival modeling.
Cox proportional hazards models: Three Cox Proportional Hazards models were fit: (i) using both clinical and geometric features, (ii) using only clinical features, and (iii) using only geometric features. The binary Follow-up Status (FUS) column served as the censoring indicator, and Overall Survival (OS) in years served as the time-to-event variable. Model performance was evaluated using Harrell's C-index, where 0.5 corresponds to random prediction and 1.0 to perfect discrimination.
Bootstrap validation for optimism correction: Rather than relying on naive training-set C-index values (which overestimate real-world performance), the authors applied a bootstrap-based optimism correction procedure. 1000 bootstrap replicates were sampled with replacement from the original dataset. For each replicate, the model was fit on the bootstrap sample and evaluated on both the bootstrap sample and the original dataset. The average difference in performance between these two evaluations - the "optimism" - was subtracted from the original C-index to yield an optimism-corrected estimate. Confidence intervals were derived using the non-parametric percentile bootstrap method with 1000 replicates.
Results: The optimism-corrected C-index for the model using only geometric features was 0.635 (95% CI: 0.574, 0.691), significantly above the 0.5 random baseline and demonstrating that nuclear shape alone carries meaningful prognostic information. The clinical-only model achieved a C-index of 0.674 (95% CI: 0.602, 0.737), and the combined clinical plus geometric model achieved the highest performance at 0.700 (95% CI: 0.651, 0.744). The improvement from adding geometric features to clinical features was not statistically significant in this cohort, but the geometric-only signal is notable given that morphologic classification has historically been challenged.
DLBCL-Morph is organized into three top-level folders - TMA, Patches, and Cells - alongside clinical data files. The clinical data is stored in two formats: clinical_data.xlsx (raw values) and clinical_data_cleaned.csv (all 209 patients with outcomes recorded, with categorical variables converted to numeric codes, e.g., neg/pos/no data converted to 0/1/NaN for IHC markers). Each patient has a unique identifier that links records across all files.
TMA folder: Contains 42 digitally-scanned TMAs organized into subfolders by stain. TMA filenames are consistent across stains, so a shared TMA_id links the same patients' cores across H&E and the five IHC stains. The core.csv file maps TMA_id and core row/column position (starting from 0,0 at the upper-left) to patient_id. Each patient has two cores (duplicates). The annotations.csv file contains ROI coordinates linked to patient_id, TMA_id, and stain, specifying upper-left and lower-right corner pixel positions at 40x magnification.
Patches and Cells folders: The Patches folder contains subfolders by stain, then by patient_id, with patch PNG files named by patch_id. 195 patients have at least one patch from at least one stained TMA. The Cells folder contains subfolders by patient_id, then patch_id, with binary NPY segmentation arrays named by cell number (non-consecutive, since non-tumor cells are discarded). All geometric features are centralized in cell_shapes.csv, linkable to NPY images through patch_id and cell number.
Jupyter notebooks: The authors provide three publicly available notebooks on GitHub (github.com/stanfordmlgroup/DLBCL-Morph) covering: (1) reproduction of survival regression results from the Technical Validation section using clinical and geometric variables, (2) visualization and computation of geometric features for any segmented nucleus in the dataset, and (3) extraction of patches uniformly from inside any ROI in the dataset, useful for researchers working directly with SVS files. All geometric feature computation code is also available in the same repository.
DLBCL-Morph addresses a persistent gap in computational pathology research: the availability of annotated, multimodal digital pathology datasets with accompanying clinical outcome data for lymphoma. Most prior AI lymphoma studies have relied on proprietary institutional datasets that could not be shared, preventing reproducibility and independent validation. By releasing a publicly available dataset of 42 digitally-scanned TMAs with expert ROI annotations, per-nucleus binary segmentation images, computed geometric features, and survival data, this work substantially lowers the barrier for other research groups to develop and validate morphologic AI models for DLBCL.
The case for morphologic features: The finding that geometric features computed from H&E-stained tumor nuclei achieve a C-index of 0.635 - without any clinical variables - is notable given decades of failure to identify morphologic prognostic factors in DLBCL by human pathologists. This suggests that the prognostic signal exists within the quantitative geometry of tumor nuclei but is too subtle or multidimensional for visual assessment. The automated pipeline, combining HoVer-Net segmentation with structured geometric feature computation, operationalizes this analysis at scale across hundreds of patients and thousands of nuclei per patient.
Immunohistochemical TMAs: Beyond H&E morphology, the inclusion of five IHC-stained TMAs for CD10, BCL6, MUM1, BCL2, and MYC enables future work to extract protein expression patterns computationally. Automated quantification of IHC staining intensity and spatial distribution could provide continuous-valued biomarkers that improve on the binary positive/negative thresholds applied in current clinical algorithms like the Hans COO classifier.
Limitations and external validation: The authors explicitly note that the survival regression findings should be validated in external datasets and prospectively in future studies. The dataset is drawn from a single institution (Stanford), and the 170-patient effective analysis cohort is modest by deep learning standards. The geometric-only C-index of 0.635, while statistically meaningful, leaves substantial room for improvement through larger cohorts, more sophisticated modeling approaches, or integration with molecular data. The correct interpretation is that DLBCL-Morph establishes the feasibility of morphologic prognostication rather than providing a deployment-ready clinical tool.