Multi-OSCC: Multiscale & Multimodal Analysis

• Multi-OSCC is a comprehensive framework that integrates multi-magnification, multi-region, and multi-task imaging into a cohesive oral cancer analysis tool.
• The canonical dataset includes 1,325 OSCC patients with six high-resolution H&E images per patient, capturing both core and edge tissue details.
• Empirical studies show that strategic fusion methods, including simple concatenation and edge-focused imaging, significantly enhance diagnostic and prognostic performance.

Multi-OSCC most explicitly denotes a recent oral squamous cell carcinoma (OSCC) histopathology dataset that integrates multi-magnification, multi-region, and multi-task supervision, but the term is also used more broadly for computational frameworks that combine multiple images, modalities, anatomical sites, or clinical endpoints in OSCC analysis. In its canonical dataset form, Multi-OSCC comprises 1,325 OSCC patients, six high-resolution hematoxylin and eosin (H&E) images per patient at ×200, ×400, and ×1000 magnifications, and six labels spanning recurrence prediction and five diagnostic tasks (Guan et al., 22 Jul 2025). Related work extends the same multi-source logic to weighted multimodal fusion across clinical, radiological, and histopathological images (George et al., 4 Oct 2025), multicenter oral cytology (Jain et al., 11 Jun 2025), multi-site confocal laser endomicroscopy (Aubreville et al., 2017), and multi-station lymph node analysis on contrast-enhanced CT (Wu et al., 2023).

1. Terminology and scope

The most explicit use of the name appears in the dataset paper "A High Magnifications Histopathology Image Dataset for Oral Squamous Cell Carcinoma Diagnosis and Prognosis," which introduces Multi-OSCC as a histopathology image dataset for OSCC diagnosis and prognosis (Guan et al., 22 Jul 2025). A second usage appears in "Multi-Modal Oral Cancer Detection Using Weighted Ensemble Convolutional Neural Networks," where Multi-OSCC denotes a multimodal weighted-ensemble framework that integrates clinical intraoral photographs, radiological imaging, and histopathology for binary OSCC detection (George et al., 4 Oct 2025).

The literature also contains technically adjacent resources that are described as aligned with the same research intent without always using the exact name. In one synthesis of the oral cytology resource "A Cytology Dataset for Early Detection of Oral Squamous Cell Carcinoma," “Multi-OSCC” refers to that multicenter dataset, while the paper itself does not explicitly brand it as “Multi-OSCC” (Jain et al., 11 Jun 2025). Likewise, the OCDC histology segmentation dataset is described as directly usable in multi-image pipelines because it consists of patch-level images sampled from whole-slide images (WSIs) and supports aggregation across many patches per WSI or patient (Santos et al., 2023).

This suggests that, within OSCC AI research, “Multi-OSCC” functions less as a single standardized artifact than as a family of multi-image, multi-scale, multi-task, multimodal, or multicenter strategies for oral cancer modeling.

2. Canonical Multi-OSCC dataset

The dataset introduced under the name Multi-OSCC was created to address two limitations in publicly available OSCC datasets: small cohorts and a narrow focus on either diagnostic or prognostic targets (Guan et al., 22 Jul 2025). It contains 1,325 unique patients with surgically treated OSCC from Sun Yat-sen Memorial Hospital, Sun Yat-sen University, collected between 2015 and 2022. Inclusion criteria were histopathologically confirmed squamous cell carcinoma, surgical intervention, and at least 2-year clinical follow-up. Each patient is represented by six high-resolution H&E images: two tissue sections, core and edge or border of the lesion, each imaged at ×200, ×400, and ×1000 magnification. Native image resolution is 2592 × 1944 pixels, and the staining modality is H&E (Guan et al., 22 Jul 2025).

The dataset is organized around six patient-level tasks, one prognostic and five diagnostic.

Task	Type	Classes
REC	Prognosis	2-class recurrence prediction within two years post-surgery
LNM	Diagnosis	2-class lymph node metastasis in head/neck
TD	Diagnosis	3-class tumor differentiation
TI	Diagnosis	2-class local tumor invasion of surrounding tissues
CE	Diagnosis	2-class vascular invasion
PI	Diagnosis	2-class perineural invasion

Ground truth is derived from hospital electronic medical records and pathology reports, with recurrence labels based on at least 2-year follow-up (Guan et al., 22 Jul 2025). The patient-level split is 925 for training, 200 for validation, and 200 for testing. To balance class distributions across tasks, patients were grouped by identical 6-task label tuples and randomly allocated within groups to train, validation, and test subsets. This prevents leakage and preserves balanced class proportions across tasks (Guan et al., 22 Jul 2025).

The dataset design is notable for combining region and magnification explicitly. Core sections emphasize keratin pearls and surrounding tumor cells, whereas edge sections capture cancerous tissue and adjacent structures; increasing magnification exposes finer cellular detail. This coupling of core versus edge morphology with ×200, ×400, and ×1000 views makes the dataset structurally suited to controlled studies of multi-image fusion and cross-task representation sharing (Guan et al., 22 Jul 2025).

3. Baselines, fusion strategies, and empirical findings

The Multi-OSCC benchmark evaluates a broad set of visual encoders, including ImageNet-pretrained ResNet-50, DenseNet-121, ViT-Base, ViT-Small, and Swin Transformer V2 Base, as well as pathology-pretrained transformers: ViT-Small with PathoBench, ViT-Base with CONCH, and ViT-Base Hibou-B (Guan et al., 22 Jul 2025). Images are resized from 2592 × 1944 to 512 × 512 for training. Preprocessing includes z-score normalization, random cropping, and random rotation; probabilistic augmentations at $p=0.5$ include contrast, sharpness, horizontal and vertical flips, and additional contrast adjustments. Stain normalization is evaluated with Reinhard, Vahadane, and Macenko methods. Class imbalance is mitigated with SMOTE (Guan et al., 22 Jul 2025).

Multi-image fusion is performed after a shared encoder across the six images per patient. Fusion modules include concatenation, Low-Rank Multimodal Fusion, Gated Fusion, and a 2-layer Transformer-based fusion module. Classification heads are per-task multilayer perceptrons with hidden sizes $768 \rightarrow 256 \rightarrow 128 \rightarrow 64$, ReLU and LayerNorm after each hidden layer, and dropout $=0.5$ before output (Guan et al., 22 Jul 2025). The multi-task formulation uses hard parameter sharing with task-specific heads and a summed objective $L = \sum_{t=1}^{T} w_t L_t$, with Sum Loss, GradNorm, and Pareto MTL evaluated as weighting strategies (Guan et al., 22 Jul 2025).

Several benchmark findings define the current empirical picture. Hibou-B ViT-Base achieves the top recurrence AUC of approximately 94.72%, with a 95% bootstrap confidence interval of 89.78–97.35, while CONCH ViT-Base achieves the top tumor differentiation AUC of approximately 81.23%, with a confidence interval of 74.15–87.58. All tasks exceed 70% AUC in top configurations (Guan et al., 22 Jul 2025). Among fusion strategies, simple concatenation delivers the best mean AUC across tasks, 75.00%, outperforming the heavier Transformer and Low-Rank Multimodal Fusion modules in mean performance (Guan et al., 22 Jul 2025).

Region choice also matters. Core-only images yield a mean AUC of 69.59%, edge-only images 74.73%, and combined core plus edge images 75.00% (Guan et al., 22 Jul 2025). Edge-only performance exceeding core-only on multiple tasks, particularly LNM, CE, and PI, is consistent with the interpretation that border histology carries complementary information for invasiveness-related endpoints.

Stain normalization is task-dependent. Diagnostic tasks generally benefit, with CE rising to 75.52% under Reinhard normalization, whereas recurrence prediction is sensitive to original color and degrades under normalization, for example from 93.13% on original images to approximately 90.53% with Reinhard (Guan et al., 22 Jul 2025). Multi-task learning remains difficult: in the reported benchmark, multi-task learning with GradNorm and Reinhard normalization is lower than single-task baselines by approximately 3.34% mean AUC across the six tasks (Guan et al., 22 Jul 2025).

4. Adjacent histopathology resources for multi-image OSCC analysis

A key complementary resource is the Oral Cavity-Derived Cancer dataset, or OCDC, introduced for pixel-level tumor-region segmentation in H&E histopathology (Santos et al., 2023). OCDC contains 1,020 histological image patches of size 640 × 640 pixels, digitized at 20× magnification with a pixel-level resolution of $0.5025\ \mu m \times 0.5025\ \mu m$, stored in RGB color space, and paired with binary tumor-versus-background masks (Santos et al., 2023). Tissue specimens come from 15 OSCC-affected patients collected between 2006 and 2013 from the Department of Oral and Maxillofacial Pathology archives at the Federal University of Uberlândia, and whole-slide images were acquired with the Slide Scanner Aperio AT2 at 20× magnification (Santos et al., 2023).

Tumor regions were hand-annotated by a specialist using GIMP with a pen on a touch screen monitor and were fully validated by a pathologist (Santos et al., 2023). The dataset includes both tumor and normal regions, with examples of serous salivary gland, connective tissue, mucous salivary gland, striated muscle, keratinized epithelial tissue, and oral mucosa, but these normal tissue types are not provided as separate class labels; labels are binary tumor masks (Santos et al., 2023). The split described in the paper is 840 training images and 180 test images, with no separate validation set and no reported baseline performance metrics (Santos et al., 2023).

OCDC is relevant to Multi-OSCC research because the patches are randomly extracted from WSIs and therefore support WSI- or patient-level aggregation. This suggests direct use in multi-image pipelines that stitch patch-level segmentations, aggregate tumor-presence predictions across tiles, or combine segmentation with auxiliary detection objectives derived from the binary masks (Santos et al., 2023). Its limitations are equally clear: single-center acquisition, 15 patients, a single scanner model, binary labels only, and fixed 20× magnification (Santos et al., 2023).

5. Multimodal and cytology extensions

In the multimodal detection framework also called Multi-OSCC, three modality-specific DenseNet-121 models are trained via transfer learning on clinical, radiological, and histopathological images, each with a tailored classification head and modality-specific preprocessing and augmentation, and then fused by a validation-weighted ensemble (George et al., 4 Oct 2025). Validation accuracy is 100% for the radiological model, 95.12% for the histopathological model, and 63.10% for the clinical model; the corresponding ensemble weights are $\alpha_{clinical}=24.43\%$, $\alpha_{radiological}=38.72\%$, and $\alpha_{histopathology}=36.83\%$. On a multimodal validation set of 55 samples, the ensemble reaches 84.58% accuracy (George et al., 4 Oct 2025). The paper explicitly notes that patient-level splits and matched patient triads were not available, which introduces leakage risk and weakens modality coherence at the patient level (George et al., 4 Oct 2025).

The oral cytology line of work extends Multi-OSCC into minimally invasive screening. The multicenter cytology dataset introduced in 2025 is described as the first large, multicenter oral cytology dataset for OSCC and includes 234 patients, 368 WSIs, and dual staining with PAP and MGG collected from ten tertiary medical centers in India (Jain et al., 11 Jun 2025). The slides were digitized at 40× with 3DHISTECH scanners, stored in MRXS format, and annotated at nucleus level using QuPath-compatible polygons. The annotated subset contains 39,246 nuclei across four categories: 19,888 in Category I, 6,175 in Category II, 6,022 in Category III, and 7,161 in Category IV (Jain et al., 11 Jun 2025). Center-based train/validation/test splitting is used to assess generalization, and HoverNet reports AJI 0.752, DQ 0.758, SQ 0.743, and PQ 0.542 on validation for nucleus instance segmentation (Jain et al., 11 Jun 2025).

Weakly supervised patient-level diagnosis has subsequently been built on that resource. RAA-MIL introduces patient-level weak labels—Healthy, Benign, Oral Potentially Malignant Disorders, and OSCC—for a 162-patient subset, represents each case as a bag of 2048 × 2048 cytology patches resized to 224 × 224, tokenizes each patch into a 14 × 14 grid using a frozen DINO-pretrained ViT-S/16 encoder, and refines local token features with Region-Affinity Attention before gated MIL pooling (Mukherjee et al., 15 Nov 2025). On a held-out test set of 33 patients, RAA-MIL achieves 72.73% accuracy, weighted F1-score 0.6970, and weighted PR-AUC 0.7969, improving over a vanilla MIL baseline; for the OSCC class specifically, PR-AUC improves from 0.5437 to 0.7667 (Mukherjee et al., 15 Nov 2025).

Taken together, these multimodal and cytology studies move Multi-OSCC beyond conventional H&E microscopy toward integrated triage systems, weakly supervised slide diagnosis, and cross-center robustness.

6. Multi-site imaging, anatomical context, and ambiguity of the label

Multi-OSCC methods are not limited to static pathology. In confocal laser endomicroscopy, a single site-agnostic classifier was trained across four oral cavity locations—inner lower labium, upper alveolar ridge, hard palate, and lesion site—using 7,894 good-quality CLE images from 12 patients after artifact exclusion (Aubreville et al., 2017). The patch-based CNN with patch-probability fusion achieved 88.3% accuracy, 86.6% sensitivity, 90.0% specificity, and AUC 0.955 under leave-one-patient-out cross-validation, while also producing dense probability maps for lesion boundary visualization (Aubreville et al., 2017). This is a multi-site OSCC paradigm rather than a multi-magnification one, but it follows the same logic of pooling heterogeneous local views into a single patient-robust detector.

A similar expansion toward anatomical context appears in metastatic lymph node detection on contrast-enhanced CT. A two-stage detector is augmented by a lymph node station classification branch with a graph convolutional network over 12 station patches per patient, integrating candidate appearance, GCN-refined station features, and candidate-to-station distances (Wu et al., 2023). On 114 OSCC CT scans, the full model reports mFROC $63.32\% \pm 1.04$, maxF1 $0.5571 \pm 0.0051$, and AUC $768 \rightarrow 256 \rightarrow 128 \rightarrow 64$0, with particular gains for small nodes and a station-classification AUC increase from 0.7314 to 0.7908 when GCN is used (Wu et al., 2023). This is a multi-region formulation in which station-level risk serves as structured context for candidate-level classification.

The field also includes portable multimodal optical screening. The smartphone-based MAF-IS platform combines autofluorescence imaging with fluorescence spectroscopy and evaluates 68 sites from 31 patients, reporting SVM-based sensitivity 88.89%, specificity 100%, accuracy 95%, and ROC AUC approximately 0.95 for OSCC versus normal classification (Thapa et al., 2022). Although not labeled Multi-OSCC in the same way as the histopathology dataset or DenseNet ensemble, it is another example of fusing complementary views of oral cancer.

A persistent source of confusion is the acronym itself. Outside oral oncology, OSCC also denotes Object State Change Classification in the Ego4D egocentric video benchmark, where models are trained for binary clip-level change versus no-change classification and report test accuracies such as 73.7% for EgoVLP and 79.6% for CSN+VideoMAE fusion (Lin et al., 2022, Zheng et al., 2022). In theoretical computer science, OSCC further denotes Online Selection with Convex Costs, an online profit-maximization problem solved by threshold policies with competitive-ratio guarantees (Tan et al., 2023). In encyclopedia usage, therefore, Multi-OSCC requires domain qualification: in biomedical AI it refers to multi-source modeling for oral squamous cell carcinoma, whereas in unrelated literatures the same acronym belongs to different problem classes.

Across the oncology literature, the main limitations recur with notable consistency: single-center acquisition for the canonical histopathology dataset, strong class imbalance for some labels such as CE and PI, absence of inter-rater agreement metrics in several resources, modest multimodal validation size, and incomplete patient-level cross-modality matching (Guan et al., 22 Jul 2025, George et al., 4 Oct 2025, Santos et al., 2023). A plausible implication is that future Multi-OSCC work will be defined less by any single benchmark and more by whether it can preserve the current gains of multi-image and multi-context modeling under external validation, center shift, and clinically coherent patient-level study design.