Primary intestinal T-cell lymphomas (PITLs) are a rare group of hematologic malignancies presenting in the gastrointestinal tract. Under the 2017 WHO classification of lymphoid neoplasms, PITLs include enteropathy-associated T-cell lymphoma (EATL), monomorphic epitheliotropic intestinal T-cell lymphoma (MEITL), intestinal T-cell lymphoma not otherwise specified (ITCL-NOS), and indolent T-cell lymphoproliferative disorder of the gastrointestinal tract. Of these, EATL carries a distinct genetic predisposition linked to celiac disease and is relatively straightforward to identify clinically. The real diagnostic challenge lies in distinguishing MEITL from ITCL-NOS, two entities that can appear morphologically similar despite representing different diseases with different underlying biology.
MEITL characteristics: MEITL (formerly called type II EATL) is defined by monomorphic small- to medium-sized tumor cells growing in an epitheliotropic pattern and expressing CD8, CD56, and cytotoxic markers. The key word is "monomorphic" - the cells are uniform in shape and size. ITCL-NOS, by contrast, is a diagnosis of exclusion comprising a heterogeneous group of peripheral T-cell lymphomas (PTCLs) that do not meet criteria for any specific PITL subtype. ITCL-NOS cells are typically medium to large-sized and pleomorphic, meaning they vary considerably in shape and nuclear appearance.
The subjectivity problem: Distinguishing MEITL from ITCL-NOS relies heavily on morphological evaluation of tumor cell nuclei, which is inherently subjective. Even experienced hematopathologists can disagree on cases that fall in a gray zone between the two entities. Cases with MEITL immunophenotype but more pleomorphic nuclear features, or MEITL-like cellular morphology with atypical immunophenotype, create genuine diagnostic uncertainty. This subjectivity has direct consequences: MEITL and ITCL-NOS have different clinical behaviors and the field increasingly needs objective, quantitative criteria for distinguishing them.
This study, from researchers at aetherAI and multiple Taiwanese hospitals, presents a machine learning workflow using quantitative nuclear morphometrics to classify PITL cases objectively. The approach is explicitly designed to be human-interpretable, generating clinically meaningful features that pathologists can understand and validate rather than functioning as an opaque black box.
The researchers collected 40 PITL patients from 17 hospitals across Taiwan, comprising 26 MEITL cases, 10 ITCL-NOS cases, and 4 borderline cases with inconsistent immunophenotypic and morphological features. All diagnoses were confirmed by a senior hematopathologist (S.S.C.) using 2017 WHO criteria. All 40 specimens were surgical resections, and all cases tested negative for Epstein-Barr virus by in situ hybridization. This multi-institutional cohort, spanning 17 centers, helps mitigate the single-center bias common in AI pathology studies, though the total case count remains modest given the rarity of PITL.
Whole-slide image preparation: Each patient contributed one hematoxylin-and-eosin stained section, digitized using the Pannoramic 250 Flash scanner (3DHISTECH) at 40x objective. The resulting 40 whole-slide images (WSIs) had an average size of 175,274 x 407,126 pixels. Before scanning, a senior hematopathologist manually selected the most representative, high-quality tissue region on each slide to avoid artifacts from out-of-focus scanning, tissue folding, or air bubbles - a critical quality control step that ensures model training occurs on diagnostically informative tissue.
Two-tier dataset structure: The study used two distinct datasets for two distinct tasks. For training the lymphocyte detection model, 33 regions of interest (ROIs, each 768 x 768 pixels, corresponding to 115 x 115 micrometers) were extracted from 19 cases and split 8:1:1 into training, validation, and test sets at the ROI level, with test ROIs drawn from WSIs not used in training. For the classification model, 400 high-power fields (HPFs, each 2304 x 2304 pixels, corresponding to 346 x 346 micrometers) were sampled from all 40 cases, with 10 HPFs randomly selected per case. A 3-fold cross-validation framework was applied at the case level, ensuring images from any individual patient appeared in only one fold - a critical precaution against data leakage that many published studies overlook.
Borderline case handling: The 4 borderline cases presented a nuanced challenge. Phenotypically, they all expressed CD8, CD56, and cytotoxic markers consistent with MEITL, but their tumor cells were larger than typical MEITL. The researchers excluded these from training and testing and used them purely as a reference set to evaluate how the model handles diagnostically ambiguous specimens.
The first computational stage was building a model to detect and precisely segment the boundaries of individual lymphoma cell nuclei in each high-power field. For this task, the researchers employed HTC-RCNN (Hybrid Task Cascade Regional Proposal Convolutional Network) with ResNet50 as the backbone. HTC-RCNN is a sophisticated instance segmentation architecture that jointly optimizes both object detection (finding where nuclei are) and instance segmentation (delineating exact pixel boundaries for each nucleus). Training used stochastic gradient descent (SGD) with a learning rate of 0.001 and batch size of 16 on a single NVIDIA V100 GPU.
Ground truth annotation: For supervised training, the boundary contours of lymphocytes in each ROI were manually annotated by two pathologists (C.H.L. and S.S.C.), providing high-quality ground truth labels. Detection performance was quantified using average precision (AP), computed from the precision-recall curve with an intersection-over-union (IoU) threshold of 0.5. On the test set, the model achieved AP of 0.881, with precision (positive predictive value) of 0.911 and recall (sensitivity) of 0.868 - meaning roughly 9 in 10 detected cells were real, and about 87% of true cells were found.
Seven morphometric attributes: Once nuclei were segmented, seven numerical attributes were computed for each individual nucleus: (1) ratio of axis length - the ratio of the longest to second-longest nuclear axis, capturing nuclear elongation; (2) circularity - the ratio of overlap between a concentric circle and cell size; (3) entropy - a measure of pixel randomness within the cell; (4) nuclear area - total pixels within the nucleus boundary; (5) irregularity - variance of distance from nucleus center to each boundary vertex; (6) perimeter - total pixels along the cell boundary; and (7) orientation - the angle of the longest nuclear axis. These attributes were chosen because they capture the key aspects of nuclear shape and texture that pathologists use when visually assessing pleomorphism.
Statistical aggregation: For each of the 7 attributes in each HPF, four moment statistics were computed over all nuclei in that field: mean, variance, skewness, and kurtosis. This yields 28 features per HPF (7 attributes x 4 moments), which were then averaged across all 10 HPFs per case to produce a 28-dimensional feature vector representing the morphological profile of each lymphoma case. The entire morphology extraction pipeline ran at an average of 2.22 seconds per HPF, making it feasible for potential integration into clinical workflows.
With 28 morphometric features per case established, the researchers trained an XGBoost (Extreme Gradient Boosting) classifier to distinguish MEITL from ITCL-NOS. XGBoost is a decision-tree-based ensemble method that builds many sequential trees, each correcting the errors of the previous one, and is well-suited to tabular feature data. Feature selection within XGBoost used Gini-based weighting applied iteratively to reduce redundant features and minimize binary cross-entropy loss. Hyperparameter optimization was performed via random search. The model was evaluated at the case level, aggregating HPF-level predictions across the 10 HPFs per patient.
Two XGBoost model variants: The researchers trained two versions of the XGBoost model. Model XGB-1 used only the 28 morphological feature scores as inputs. Model XGB-2 added immunohistochemical (IHC) phenotype data - specifically the positive/negative status for CD8 and CD56 expression - on top of the morphological features. This design tests whether adding molecular marker information improves classification beyond morphology alone, which has direct practical implications: IHC requires additional staining, time, and cost, so if morphology alone is sufficient, the workflow is simpler.
CNN baseline comparison: To benchmark the XGBoost approach against an end-to-end deep learning strategy, a separate CNN model was trained using ResNet50, taking entire HPF images directly as inputs (without any morphometric feature extraction as an intermediate step). This CNN used the AdamW optimizer at learning rate 0.001 with batch size 16 on an Nvidia Quadro RTX 8000. Extensive data augmentation was applied including random horizontal and vertical flipping, random translation, rotation, scaling, color jittering (brightness, contrast, saturation, hue), Gaussian blurring, and random cropping. Early stopping halted training when validation loss did not improve for 10 consecutive epochs.
Statistical comparison methodology: Classification performance was assessed using AUC under the receiver operating characteristic curve with 3-fold cross-validation at the case level. Differences between model AUCs were formally tested using Delong's test, a paired comparison method for AUC values with a two-tailed significance level of 0.05. A general linear model (GLM) with repeated measurements was also applied to test which individual morphological features showed statistically significant differences between MEITL and ITCL-NOS cases, with F-statistics computed for each of the 28 features.
The morphology-only XGBoost model (XGB-1) achieved an AUC of 0.966 (95% CI: 0.949-0.984) for classifying MEITL versus ITCL-NOS. Adding IHC phenotype data (XGB-2) yielded an AUC of 0.955 (95% CI: 0.935-0.975), which was not significantly different from XGB-1 by Delong's test (p = 0.412). This is a particularly important finding: immunohistochemical markers CD8 and CD56 - which are part of the standard diagnostic workup for PITL - added no meaningful discriminative power on top of morphological features alone. In other words, the shape and texture of the tumor nuclei contain essentially all the information needed for classification.
CNN performance and the key comparison: The end-to-end CNN trained directly on HPF images achieved an AUC of 0.820 (95% CI: 0.734-0.906). This was significantly inferior to both XGBoost models: p = 0.003 for XGB-1 vs. CNN and p = 0.001 for XGB-2 vs. CNN. The magnitude of the difference is notable - approximately 0.14-0.15 AUC units separating the approaches. The XGBoost model operating on expert-defined morphometric features substantially outperformed the data-driven CNN despite the CNN having access to all raw image information. This result challenges the assumption that end-to-end deep learning always extracts more information than handcrafted features.
Case-level concordance: There was high concordance between XGBoost predictions and the diagnoses of the senior hematopathologist, with three exceptions: Case T20 was diagnosed ITCL-NOS by the pathologist but predicted as MEITL by the model, while Cases T02 and T70 were diagnosed MEITL but predicted as more likely ITCL-NOS. These discordant cases likely represent genuine gray-zone specimens where even expert pathologists face uncertainty, and the model may be capturing subtle morphological signals that challenge intuitive visual assessment.
Why XGBoost beat the CNN: The authors reason that CNNs suffer from overfitting when training data is limited. With only 36 cases in the classification training folds, the ultra-high dimensionality of raw image features overwhelms a neural network's capacity to generalize. XGBoost with 28 human-defined features avoids this curse of dimensionality, learning from a compact representation that captures the biologically meaningful variance. The CNN's AUC of 0.82 is not poor, but it represents the cost of insufficient training data for an architecture that needs much more.
The XGBoost model's feature importance, measured by "gain" (how much each feature reduces loss when used as a split point across all trees), revealed that the top-ranked discriminative features were variance in nuclear perimeter, variance in nuclear area, and mean nuclear irregularity. When IHC markers were included in XGB-2, CD56 and CD8 expression ranked above most morphological features - but still below variance in perimeter - confirming that perimeter variability is the single most informative signal in this classification task. This aligns with biological intuition: MEITL cells have small, round, regular nuclei (low perimeter variance), while ITCL-NOS cells are large and irregular (high perimeter variance).
Statistical testing with GLM: The general linear model confirmed significant differences between MEITL and ITCL-NOS across multiple morphological features. Among the 28 computed feature scores, five of seven attributes showed significant differences in variance (p less than 0.01), and five showed significant differences in mean. For nuclear area, ITCL-NOS cases had larger mean area (45.84 +/- 6.93 square micrometers) compared to MEITL (38.13 +/- 7.25 square micrometers, p less than 0.001). For irregularity, ITCL-NOS showed substantially higher mean irregularity (1.55 +/- 0.30) than MEITL (0.94 +/- 0.22, p less than 0.001). Features related to perimeter and irregularity were the only ones showing significant differences in both skewness and kurtosis, indicating not just different average values but different distributional shapes across the cell population.
Two-dimensional visualization: Plotting variance in nuclear perimeter against variance in nuclear irregularity produced a striking visual separation of cases. MEITL cases clustered in the lower-left corner (small variance in both features, reflecting uniform small round cells), while ITCL-NOS cases occupied the upper-right corner (large variance in both, reflecting mixed cell sizes and irregular shapes). The borderline cases fell in intermediate positions, visually confirming that the model's uncertainty about these cases has a quantitative morphological basis.
Cell orientation showed no statistically significant difference between disease subtypes, which makes biological sense - the directionality of nuclear axis is not expected to be diagnostic for lymphoma subtype. This specificity in feature importance provides important validation that the model is learning biologically relevant signals rather than confounding artifacts.
The study makes a deliberate methodological choice in favor of human-interpretable AI over end-to-end black-box models. The authors argue that CNN classifiers operating directly on images - even when accurate - offer limited clinical utility because physicians cannot interrogate how the decision was made. In contrast, the XGBoost approach provides explicit feature weights that pathologists can map onto their existing conceptual understanding of nuclear morphology. A pathologist reviewing an ambiguous case can look at the computed irregularity variance and perimeter variance scores and connect those numbers to what they see microscopically, allowing the AI output to inform rather than replace clinical judgment.
Comparison to QuPath and traditional methods: The authors benchmarked their deep learning segmentation against QuPath (an open-source digital pathology tool widely used in research) using default settings. QuPath frequently produced fragile contours and false positives, particularly for crowded cells, cells with vesicular nuclei, and cells with atypical chromatin patterns. The HTC-RCNN model showed more robust segmentation with fewer fragile contours, though it introduced more false negatives. Traditional methods like QuPath require careful manual adjustment of hyperparameters (background radius, nucleus radius, local intensity threshold) for each image, while the deep learning model applies a consistent learned approach across all fields.
Context within the broader AI pathology literature: Prior AI studies in lymphoma have achieved high AUC values (over 0.95) for diffuse large B-cell lymphoma and follicular lymphoma classification using CNNs, but these tasks exploited architectural patterns (follicular density, follicle size) rather than single-cell morphology. PITL grows in a diffuse pattern without the distinctive architectural hallmarks that make those tasks tractable for CNNs. The shift to single-cell morphometrics is precisely what enables classification of PITL despite the diffuse growth pattern and despite limited sample sizes.
Borderline case insights: The 1:1 prediction split of the four borderline cases (two predicted as MEITL, two as ITCL-NOS) suggests that even within this diagnostically ambiguous group, there may be genuine morphological differences. The authors speculate that the strict "monomorphic" criterion in MEITL's name may not capture the full spectrum of the disease, and that some degree of nuclear pleomorphism may be compatible with what currently defines MEITL. The underlying genetic architecture of these borderline cases was under active investigation at the time of publication and may ultimately provide ground truth for resolving their classification.
Sample size constraints: The most significant limitation is the small dataset: 36 cases in the classification model (26 MEITL, 10 ITCL-NOS), collected from patients in Taiwan. PITL is genuinely rare, making large-scale data collection difficult. The substantial class imbalance (26 vs. 10) further complicates interpretation. The authors address this with 3-fold case-level cross-validation, but external validation on an independent cohort - ideally from multiple countries - is needed before the model could be considered clinically deployable.
Non-lymphoid cell contamination: The lymphocyte detection model was trained exclusively on lymphocyte annotations. As a result, cells from other tissue types - particularly vascular endothelial cells - may be incorrectly included as lymphocytes. These falsely included non-lymphoid cells could distort morphometric feature estimates, particularly in fields with abundant stromal elements. Including annotations for multiple cell types in future training would help the model discriminate lymphoma cells from other tissue components more reliably.
Potential for broader application: The authors explicitly frame this workflow as a template applicable beyond PITL. The same pipeline - deep learning segmentation followed by XGBoost classification on morphometric features - could be applied to other challenging lymphoma differential diagnoses, such as distinguishing classic versus pleomorphic cytomorphological variants of mantle-cell lymphoma, or classifying low-grade versus high-grade follicular lymphoma. The workflow's reliance on H&E staining only, without requiring IHC, makes it particularly attractive for resource-limited settings.
Clinical integration pathway: In practice, the authors envision pathologists using the AI-derived morphological indexes alongside IHC results, particularly for borderline cases. Rather than replacing the pathologist's judgment, the system provides an objective, quantitative second opinion on nuclear morphology. Automated scoring of morphological features like these could also enable analyses at a scale far beyond what is feasible with manual visual scoring - examining thousands of cells per case rather than the few hundred accessible during standard microscopy. Future work connecting morphological features to patient outcomes and genetic subtypes would further establish the clinical utility of this approach.