Sarcomas are a clinically significant group of mesenchymal cancers that are particularly challenging to diagnose because there are over 50 recognized subtypes, each with distinct biological behavior, prognosis, and treatment implications. In children, the two major categories are rhabdomyosarcomas (RMS) and non-rhabdomyosarcoma soft tissue sarcomas (NRSTS), with alveolar RMS carrying a substantially worse prognosis than the more common embryonal subtype. Accurate subtyping is critical because misclassification directly harms prognosis and treatment planning, yet formal diagnostic rates in resource-limited settings remain far below those at specialized pediatric cancer centers.
The diagnostic bottleneck: Standard-of-care sarcoma diagnosis at most major institutions depends on a combination of histopathologic review of hematoxylin and eosin (H&E)-stained slides, immunohistochemistry (IHC) panels for markers such as myogenin and MyoD1, and confirmatory molecular or genetic testing including FISH for fusion gene status. A published Brazilian sarcoma reference center study found high rates of histopathologic discordance on referral cases, and international literature consistently documents interobserver variability even among expert pathologists. These challenges are compounded by the fact that specialized sarcoma pathology expertise is highly concentrated in academic centers in high-income countries.
The AI opportunity: Deep learning models trained on digitized whole slide images (WSI) have the potential to provide standardized, reproducible classification that can be deployed remotely via web-based interfaces without requiring on-site expertise. However, a persistent barrier to building generalizable models for pediatric sarcoma has been the scarcity of large, annotated, multi-institutional datasets. Single-center training sets are susceptible to overfitting on institution-specific artifacts such as scanner hardware, staining protocols, and tissue preparation methods, reducing performance when deployed elsewhere.
This preprint from 2025 describes a collaborative effort across St. Jude Children's Research Hospital, Massachusetts General Hospital, Yale New Haven Children's Hospital, and the Children's Oncology Group (COG) to address this gap by compiling the largest known curated multi-institutional pediatric sarcoma histology dataset, harmonizing it through a standardized processing pipeline, and training deep learning classifiers that achieve strong performance across multiple clinically relevant classification tasks.
The dataset assembled for this study comprises 867 whole slide images from 722 patients, contributed by three major academic medical centers and the Children's Oncology Group. The collection spans 10 pediatric sarcoma subtypes with heterogeneity in sarcoma subtype, anatomical tumor location, patient sex, and patient race and ethnicity. Ground truth labels were derived from the institutional pathological diagnosis provided by each contributing center, supplemented by limited associated clinical metadata including patient age at sample collection, sex, race and ethnicity, primary versus metastatic site, and molecular signature summaries where available.
Image harmonization pipeline: Each medical center employs unique tissue extraction procedures, staining protocols, scanner hardware, and image file formats. To unify these diverse inputs and reduce site-specific batch effects, the team applied an established processing pipeline called STQ (Spatial Transcriptomics Quantification) v3.0.0. The pipeline performs image format conversion to a tifffile-readable standard, focus checking using a deep learning-based DeepFocus quality control step, resolution standardization to approximately 0.25 microns per pixel (mpp), and stain normalization to reduce chromatic variability introduced by different reagent lots and scanners. Tissue artifacts such as tissue folds, mounting media blobs, air bubbles, and excessive contamination are excluded during quality control before tiling.
SAMPLER representation: Rather than training a full end-to-end deep learning classifier on raw tile sequences, the study employs the previously published SAMPLER (Statistical Aggregation Method for Pathology Image Learning and Efficient Representation) approach. For each WSI, tiles are extracted from regions of interest identified by the Automatic Object Identification (AOI) tool using DBSCAN clustering. Features are extracted per tile using a pre-trained deep learning backbone. The empirical cumulative distribution function (c.d.f.) of each feature across all tiles is then computed and summarized as 10 quantiles (5th through 95th percentile), producing a fixed-length statistical representation of the slide that is independent of tile count and order. This vector is then used to train lightweight classifiers (logistic regression or linear SVM) for downstream tasks.
Batch effect correction: To visualize and confirm the effectiveness of harmonization, the team computed top principal components of SAMPLER representations across all slides, applied the Harmony algorithm using the four contributing institutions as batch variables, computed 8 nearest-neighbor graphs, and generated UMAP embeddings. This two-dimensional view of the slide collection enabled confirmation that the batch correction successfully intermingled slides from different centers rather than clustering them by institution of origin.
A central methodological question in computational pathology is which deep learning backbone to use as a feature extractor. The study systematically benchmarked four architectures spanning a range of training strategies and data sources. InceptionV3 is a convolutional neural network (CNN) from Google's Inception family, originally pre-trained on the general-purpose ImageNet image dataset. CTransPath is a transformer-based model pre-trained on histopathology images using unsupervised contrastive learning, specifically designed for pathology classification tasks. UNI is a general-purpose self-supervised foundation model for computational pathology, pre-trained on over 100,000 H&E-stained slides using the DINOv2 self-supervised learning objective. CONCH is a visual-language foundation model for computational pathology, trained jointly on pathology images and associated text using contrastive language-image pre-training, enabling multimodal representations.
Performance hierarchy: Across the primary classification task of alveolar versus embryonal RMS, ViT-based foundation models clearly outperformed the legacy CNN architecture. UNI and CONCH both produced state-of-the-art AUROCs of 0.959 and 0.958, respectively, with no statistically significant difference between them. CTransPath, also transformer-based, significantly outperformed InceptionV3 (p < 5x10^-5) but fell significantly short of UNI and CONCH (p < 5x10^-4). This hierarchy reflects the evolution of histopathology feature extractors: general-purpose CNN architectures trained on natural images are systematically inferior to transformers pre-trained specifically on pathology slides, with larger and more diverse pre-training datasets yielding further gains.
Tile field of view (FOV) sensitivity: The choice of tile size had a substantial impact on classification performance across all backbones. Smaller tiles capture cellular-level morphological detail (individual nuclear characteristics, mitotic figures, cytoplasmic texture), while larger tiles provide tissue-level architectural context (growth patterns, stromal relationships, spatial arrangements of cell populations). The optimal tile FOV varied by backbone, indicating that different feature extractors encode spatial biology at different preferred scales. This observation directly motivated the multiscale analysis described in the next section.
These benchmarking results have practical implications beyond this study: they demonstrate that researchers building WSI classifiers for rare pediatric cancers need not default to older CNN architectures simply because they are more familiar. The open-source availability of UNI and CONCH means that the performance gains from histopathology-specific pre-training are now accessible without additional training infrastructure.
Expert pathologists do not examine a histology slide at a single magnification. They routinely zoom in and out, switching between low-power views to assess tissue architecture and high-power views to evaluate nuclear detail, mitotic activity, and cytoplasmic characteristics. Motivated by this practice, the investigators tested whether combining features from multiple tile scales could enhance classification accuracy beyond what any single scale achieved alone.
Implementation: Multiscale feature extraction was implemented by concatenating SAMPLER feature vectors derived from two different fields of view for the same backbone, effectively doubling the number of features used as input to the logistic regression classifier. For example, a "UNI 1+4" combination extracts UNI features at FOV scale 1 (fine cellular detail) and FOV scale 4 (coarser architectural context), concatenates them, and trains a single logistic regression on the combined vector. Pairwise combinations of FOVs were tested systematically for each backbone.
Consistent gains across tasks: Multiscale feature sets consistently outperformed single-scale models across classification tasks and backbones. For the alveolar versus embryonal RMS task, the best combination was CONCH FOV 1+2, which achieved mean AUROC of 0.961 over 100 iterations of 5-fold cross-validation (with a standard deviation of 0.021), compared to single-scale AUROC values that fell meaningfully below this threshold depending on which FOV was used. For the RMS versus NRSTS task, UNI 1+4 was the optimal combination, achieving mean AUC of 0.969 (standard deviation 0.026). The multiscale approach may also improve external generalizability by smoothing out artifacts that are more prominent at a single scale than across multiple scales simultaneously.
The investigators note a practical advantage of the multiscale concatenation within the SAMPLER framework: concatenation is computationally inexpensive compared to training separate models at multiple scales or building dedicated multi-scale attention architectures. Since SAMPLER representations are pre-computed and stored as feature vectors, combining scales requires only concatenating two fixed-length vectors before fitting the logistic regression, adding minimal computational overhead.
RMS versus NRSTS: Distinguishing RMS from NRSTS is a clinically important first-level classification that normally requires expert pathologist annotation supplemented by a panel of molecular tests, including positive staining for myogenin and MyoD1 to confirm skeletal muscle differentiation in RMS. The dataset contained heterogeneous representation of both categories, making this a tractable machine learning problem. Using UNI FOV 1+4, the best-performing SAMPLER classifier achieved a mean AUC of 0.969 over 100 iterations of 5-fold cross-validation, with precision, recall, and F1 scores reported across varying decision thresholds. The confusion matrix showed that most misclassifications involved cases with overlapping morphological features, which is consistent with known inter-rater variability at the RMS/NRSTS boundary.
Ewing sarcoma two-stage pipeline: Ewing sarcoma (ES) represents a particularly challenging classification problem because it is defined by the EWSR1::FLI1 fusion gene and is rare within the cohort, appearing in only 41 of 722 total cases (5.7%). Accurate identification matters clinically because targeted therapies directed against the EWS-FLI1 fusion protein are in active development, making sensitivity in identifying ES cases high priority. To address class imbalance, the investigators designed a two-stage pipeline: first, a model classifies all samples as RMS versus NRSTS (Stage 1); all samples predicted as NRSTS then feed into a second model trained to distinguish ES from other NRSTS subtypes (Stage 2). At a threshold of 0.8, this two-stage approach achieved strong sensitivity for ES detection. The Stage 1 RMS versus NRSTS ROC curve and Stage 2 Ewing versus non-Ewing ROC curve are reported separately, along with a combined confusion matrix demonstrating the pipeline's integrated performance.
Multiclass RMS subtyping: The most granular RMS classification task distinguished alveolar, embryonal, and spindle-cell RMS simultaneously. Spindle-cell RMS is both clinically and genetically distinct but visually similar to the embryonal subtype, and rare in training sets. This three-class task was significantly harder than binary classification, and performance expectations were set accordingly. The spindle-cell class, due to visual similarity to embryonal RMS and lower sample count, showed the most classification difficulty, mirroring the challenges faced by human pathologists for this subtype.
Across all tasks, SAMPLER-based logistic regression and SVM classifiers performed comparably to each other, and the choice between them did not substantially affect final AUC values. Logistic regression was preferred for its coefficient interpretability, though SAMPLER itself aggregates across tiles and does not preserve spatial context at the tile level, limiting direct biological interpretation of which tissue features drove a given prediction.
A growing trend in computational pathology is to replace statistical WSI representations with larger, more expressive neural architectures trained directly on tile sequences. Transformer encoder models that compute multi-head self-attention (MHSA) across all image tiles can, in principle, capture long-range spatial dependencies within a slide, assigning hierarchical importance scores to different tissue regions. The investigators evaluated whether this architectural complexity translated into better classification performance by benchmarking against a transformer encoder architecture inspired by prior published work (Wagner et al.).
Transformer implementation: For the MHSA transformer, each WSI was split into 224x224 tiles at 20x magnification (0.5 mpp). Each tile was encoded using CTransPath to produce a 768-dimensional feature vector. Tile vectors were then stacked into a slide-level matrix and processed by a multi-head self-attention transformer encoder that computed attention across all tile positions. A fully connected layer with global pooling produced the final binary prediction. Training used the Adam optimizer with learning rate 1e-5, binary cross-entropy loss weighted for class imbalance, and 75 epochs per fold on an NVIDIA V100 GPU (32 GB).
Results and interpretation: On the alveolar versus embryonal RMS classification task, the MHSA transformer achieved a mean AUROC of 0.922 plus or minus 0.031 across five folds, compared to CONCH 1+2 SAMPLER achieving 0.961 plus or minus 0.021. The transformer's confusion matrix showed generally strong performance on both classes, but precision, recall, and F1 scores all fell short of the SAMPLER-based method. Additionally, the transformer's saved model file was substantially larger (1.9 MB versus SAMPLER's smaller footprint), meaning greater storage and inference overhead without commensurate performance gains for this task.
The authors interpret these results in the context of data scale. Transformer architectures excel when trained on large datasets where their capacity to learn complex spatial dependencies can be fully exercised. For rare pediatric sarcomas where even a large multi-institutional effort yields fewer than 1,000 slides, the statistical SAMPLER representation provides a better bias-variance tradeoff: it trades spatial context for computational efficiency and resistance to overfitting, yielding better generalization performance at this data scale.
A fundamental limitation of the SAMPLER approach as originally designed is that it aggregates tile features across the entire WSI into a single statistical vector, discarding tile-level spatial context. This means that while the classifier can identify which slides belong to which sarcoma subtype, it cannot directly indicate which tissue regions on a given slide contributed most to that decision. To address this interpretability gap, the investigators developed a spatially resolved class probability heatmap approach that extends the SAMPLER framework to the local tile level.
Local SAMPLER construction: For each tile in a WSI, a local SAMPLER representation is constructed using that tile and its nearest spatial neighbors within a specified radius R. For example, with R=2, the tile and its 12 surrounding spatial neighbors are used to compute a local empirical c.d.f. with 10 quantiles for each imaging feature. This local representation is then fed into the same logistic regression classifier trained on whole-slide SAMPLER vectors, producing a per-tile classification probability. These tile-level probabilities are mapped back to their spatial coordinates on the WSI to generate a continuous probability heatmap overlaid on the H&E image. Increasing the radius R produces smoother heatmaps at the cost of spatial resolution.
Clinical examples: Three representative cases are described in the paper. For a synovial NRSTS case, the heatmap correctly assigns high NRSTS probability to the dominant tissue regions and high synovial sarcoma subtype probability to the same areas, confirming classifier coherence. For an alveolar RMS case, the classifier correctly assigns zero probability of NRSTS to any region and identifies the tumor region as alveolar throughout. For an Ewing sarcoma case, which exhibits some morphological features that overlap with RMS (small round blue cell morphology shared across multiple sarcoma subtypes), the heatmap reveals spatially heterogeneous regions that the classifier assessed as having RMS-like characteristics, appropriately flagging intratumoral heterogeneity rather than producing a uniformly confident NRSTS prediction.
These heatmaps serve two purposes. First, they provide pathologists with visual feedback that can confirm or interrogate model predictions, building trust through alignment with known tissue biology. Second, they create a pathway for integrating spatial transcriptomics or spatial antibody-based imaging data by defining tissue regions with distinct molecular profiles. The approach does not require retraining any additional models, since it applies the same logistic regression weights learned at the WSI level to locally computed SAMPLER representations.
Overfitting to institutional artifacts: Although the multi-institutional design and Harmony batch correction substantially reduce center-specific artifacts, the investigators acknowledge that deep learning models in pathology remain sensitive to scanner hardware, staining protocols, and tissue preparation methods. Even after stain normalization, residual scanner-specific signals can be internalized by deep learning backbones, potentially inflating internal validation performance and reducing generalizability to new centers using different equipment. This is particularly problematic for SAMPLER, which creates WSI representations by aggregating features across all tiles, meaning scanner artifacts that affect a consistent subset of tiles could systematically bias slide-level representations.
Rare subtype representation: With only 867 slides spanning 10 subtypes, many individual subtypes are represented by fewer than 100 cases. Ewing sarcoma, at 41 cases (5.7%), represents an extreme example of the class imbalance challenge pervasive in pediatric sarcoma datasets. Some subtypes in the full dataset fell below the threshold needed for reliable binary or multiclass modeling, and the paper notes that spindle-cell RMS classification suffers from both low sample counts and high visual similarity to embryonal RMS. These constraints cannot be fully overcome by algorithm design alone and require continued multicenter data aggregation.
Ground truth label quality: The ground truth labels were derived from institutional pathological diagnoses at each contributing center, which introduces heterogeneous quality depending on institutional expertise. Subtypes requiring molecular confirmation (fusion gene status in alveolar RMS, EWSR1 rearrangement in Ewing sarcoma) may have been misclassified at contributing institutions if comprehensive genetic testing was not performed, introducing label noise that is particularly problematic for models being trained to detect these molecularly-defined entities.
Future directions: The authors identify several key next steps. First, integration with molecular data (fusion gene status, cytogenetics, IHC) would enable validation that the morphological features identified by the deep learning models correspond to biologically meaningful tissue characteristics. Second, federated learning across additional pediatric sarcoma centers without requiring centralized data sharing could further scale the training dataset while preserving patient privacy. Third, prospective clinical validation studies embedding the SAMPLER-based classifiers into diagnostic workflows would establish the real-world impact on diagnostic concordance and time to accurate diagnosis, the metrics that ultimately matter for patient outcomes.