ML-SASD: Scleral Anomaly Segmentation Dataset

• The paper introduces ML-SASD as a benchmark dataset organized into Clinical, Wild, and Mix subsets to evaluate multi-label segmentation performance.
• It employs a two-stage cascaded annotation workflow providing detailed pixel-level labels for vessels, yellow/black spots, blood spots, and SSR negative masks.
• The dataset challenges models with diverse acquisition conditions and distribution shifts, evidenced by a mean Dice of 72.11% and mIoU of 58.44%.

Searching arXiv for the cited paper and closely related context papers to ground the article. arxiv_search(query="(Yu et al., 3 Jun 2026)", max_results=5, sort_by="submittedDate") Searching arXiv by identifier. {"query":"(Yu et al., 3 Jun 2026)","max_results":5,"sort_by":"submittedDate"} The Multi-label Scleral Anomaly Segmentation Dataset (ML-SASD) is a benchmark for pixel-level scleral surface anomaly segmentation introduced in the context of the TCM-inspired Artificial Intelligence Ocular Auxiliary Diagnosis System (TAO). It was presented together with the HD-DinoMoE framework as a dataset designed for complex acquisition scenarios involving clinical images, user-acquired images, multi-source distributional discrepancies, diverse anomaly morphologies, and scleral specular reflection (SSR). ML-SASD provides Clinical, Wild, and Mix settings, together with pixel-wise annotations for three anomaly categories—Vessels, Yellow and Black Spots, and Blood Spots—and an explicit SSR negative mask in the dense annotation stage (Yu et al., 3 Jun 2026).

1. Dataset definition and composition

ML-SASD is organized into three subsets. ML-SASD-Clinical, described as “Controlled/Ideal,” contains 1,075 ocular images. ML-SASD-Wild, described as “Unconstrained/Glare,” contains 1,068 images. ML-SASD-Mix, described as a “Diverse” mix of Clinical and Wild data, contains 2,143 images (Yu et al., 3 Jun 2026).

Subset	Description	Images
ML-SASD-Clinical	Controlled/Ideal	1,075
ML-SASD-Wild	Unconstrained/Glare	1,068
ML-SASD-Mix	Diverse mix of Clinical & Wild	2,143

The Clinical images were drawn from the Illustrated TCM Ocular Inspection and Syndrome Differentiation atlas and span various patients across years of clinical practice. The Wild images were captured from 18–60 year-olds using commercial mobile devices, and many subjects were recruited from laboratory personnel to enhance anomaly diversity. This pairing of atlas-derived clinical material with mobile-device captures establishes ML-SASD as a heterogeneous benchmark rather than a single-domain corpus (Yu et al., 3 Jun 2026).

A plausible implication is that the dataset was constructed to stress-test segmentation models under both controlled and uncontrolled imaging conditions. The explicit naming of “Clinical,” “Wild,” and “Mix” settings further indicates that cross-domain robustness is a central design objective.

2. Image acquisition protocol and imaging conditions

The acquisition protocol was based on a Five-Eye Positioning Method: forward, upward, downward, leftward, and rightward gaze, with the stated purpose of maximally exposing the scleral surface. In addition, manual eyelid retraction by an operator was used to expose superior and inferior fornical regions. These protocol elements are part of the dataset definition rather than post hoc preprocessing decisions (Yu et al., 3 Jun 2026).

The Clinical subset consists of high-resolution digital-camera images acquired under controlled lighting. By contrast, the Wild subset includes smartphone, tablet, and camera images acquired under varied indoor and outdoor illumination and is affected by glare and motion blur. The Clinical images are predominantly in the range of 1,400–1,500 pixels in width and 1,000–1,200 pixels in height, whereas the Wild subset spans a broad range from below 500 pixels up to approximately 3,500 pixels on one axis. All images were later resized to $1{,}024 \times 1{,}024$ for training (Yu et al., 3 Jun 2026).

This acquisition design encodes several sources of covariate shift directly into the benchmark: sensor variation, illumination variation, motion degradation, and resolution dispersion. The paper explicitly associates these conditions with “multi-source distributional discrepancies,” which situates ML-SASD within the broader class of robustness-oriented medical segmentation datasets.

3. Annotation workflow and label semantics

ML-SASD uses a two-stage, cascaded semantic annotation workflow. In Stage 1, full ocular images receive semantic masks for Periocular, Iris, and Sclera; 1,325 images were annotated at this stage. In Stage 2, a scleral ROI receives dense, multi-label masks for three anomaly classes together with a specular-reflection negative mask; 515 images in the Mix subset were annotated at this stage (Yu et al., 3 Jun 2026).

The annotations were produced by manual pixel-wise labeling by trained annotators, followed by multi-round expert review by researchers experienced in ocular image analysis. Ambiguous cases were resolved by consensus, and annotators were guided to follow visible edges without speculative expansion. No formal inter-annotator-agreement percentage was reported (Yu et al., 3 Jun 2026).

The three anomaly classes are defined morphologically. Vessels (Ve) are described as a fine, elongated, tortuous vascular network, including continuous capillary branches with variable density. Yellow and Black Spots (YBS) are plaque-like or dot-like pigmented lesions, with color ranging from pale yellow to dark black and typical occurrence often in terminal or perivascular regions. Blood Spots (BS) include red-to-dark-red hemorrhagic clots, marginal congestion, and diffuse or patch-like areas; listed variants include light-red marginal congestion, localized perilimbal hemorrhage, and diffuse dark patches (Yu et al., 3 Jun 2026).

A defining property of the dataset is its explicit multi-label rule: pixels may belong to multiple anomaly classes when visual evidence indicates overlap. SSR regions are labeled as negative samples to quantify model false positives in specular patches. This distinguishes ML-SASD from mutually exclusive semantic segmentation datasets and makes it specifically suited to overlapping lesion morphology and artifact-aware evaluation (Yu et al., 3 Jun 2026).

4. Splits, preprocessing, and evaluation conventions

The Clinical and Wild subsets are each divided roughly $8{:}1{:}1$ into train, validation, and test partitions using subject- or sequence-wise grouping. The Mix subset is formed by aggregating the corresponding splits from Clinical and Wild. No $k$-fold cross-validation is used; held-out test sets are used for final evaluation (Yu et al., 3 Jun 2026).

Preprocessing follows a fixed pipeline. Images are normalized to the ImageNet mean and standard deviation,

$$\mu=(0.485, 0.456, 0.406), \quad \sigma=(0.229, 0.224, 0.225),$$

and resized to $1{,}024 \times 1{,}024$, with bilinear interpolation for images and nearest-neighbor interpolation for masks. No additional augmentation, such as geometric or photometric augmentation, is described prior to annotation or training (Yu et al., 3 Jun 2026).

The dataset description also provides standard segmentation metrics for reference: $$\mathrm{Dice}(A,B)=\frac{2|A\cap B|}{|A|+|B|},\qquad \mathrm{IoU}(A,B)=\frac{|A\cap B|}{|A\cup B|}.$$ These formulas define the principal overlap metrics used in reporting segmentation performance. Their inclusion in the dataset documentation standardizes the evaluation vocabulary, though the benchmark’s particular emphasis on SSR false positives adds an artifact-specific dimension not captured by Dice and IoU alone (Yu et al., 3 Jun 2026).

5. Statistical structure and segmentation difficulty

At the image level, Ve, YBS, and BS each appear in 60–90% of images, and multi-class concurrence is very high, with co-occurrence greater than 0.9 for many pairs. At the pixel level in the Mix subset, background occupies approximately 90% of pixels, Ve approximately 2–4%, YBS approximately 1–3%, BS approximately 2–5%, and SSR less than 1% (Yu et al., 3 Jun 2026).

The dataset description formalizes class proportion by the class imbalance ratio

$r_c = \frac{N_c}{\sum_i N_i},$

where $N_c$ is the total pixel count of class $c$. It also reports that most anomaly regions occupy 0–5% of image area, with the lesion area distribution peaking in the small-object regime. Resolution distributions differ sharply between subsets: Clinical samples cluster in a narrow band, while Wild samples are widely scattered (Yu et al., 3 Jun 2026).

These properties jointly define the benchmark’s technical difficulty. The high background fraction implies severe foreground sparsity; the 0–5% lesion area regime indicates a small-object segmentation problem; the strong co-occurrence structure means that independent single-label assumptions are inappropriate; and the resolution disparity between Clinical and Wild partitions compounds distribution shift. This suggests that ML-SASD is not merely challenging because of lesion morphology, but because multiple failure modes—class imbalance, overlap, artifact contamination, and acquisition heterogeneity—are present simultaneously.

6. Role in scleral anomaly segmentation research

ML-SASD was introduced as a new benchmark for multi-label scleral anomaly segmentation and used to evaluate HD-DinoMoE, a class-aware hierarchical dual mixture-of-experts network. In the accompanying study, HD-DinoMoE is designed to segment Vessels, Yellow and Black Spots, and Blood Spots under complex acquisition scenarios and incorporates mechanisms specifically intended to address SSR regions, dual-backbone adaptation, and sample- and class-level imbalance. On ML-SASD-Mix, the reported performance is a mean Dice of 72.11% and a mean Intersection-over-Union of 58.44%, while maintaining favorable boundary localization and specular-region false-positive control (Yu et al., 3 Jun 2026).

The same study states that the method also shows competitive generalization on the Vessels subset of the public SBVPI dataset. Within this framing, ML-SASD functions both as an internal benchmark for the TAO system and as an externalization of the problem setting that includes clinical and user-acquired images. The benchmark is therefore tied to a concrete downstream diagnostic context while remaining evaluable as a segmentation dataset in its own right (Yu et al., 3 Jun 2026).

A common misconception would be to view ML-SASD as a conventional sclera segmentation dataset. The annotation design shows otherwise: Stage 1 includes coarse ocular structure masks, but the distinctive contribution lies in Stage 2 dense multi-label anomaly masks with an SSR negative mask. Another possible misconception is that the benchmark eliminates ambiguity through exhaustive labeling statistics; in fact, the description explicitly notes consensus-based review but no formal inter-annotator-agreement percentage. Accordingly, ML-SASD is best understood as a practically curated, expert-reviewed benchmark whose primary value lies in modeling realistic overlap, glare, and domain shift, rather than in providing exhaustive annotator-consistency quantification.

7. Limitations and methodological implications

Several limitations are explicit in the dataset description. Only 515 images in the Mix subset receive Stage 2 dense anomaly annotations, despite the larger corpus size. No formal inter-annotator-agreement percentage is reported. No $k$-fold cross-validation is performed, and no additional augmentation is described prior to training (Yu et al., 3 Jun 2026).

These facts do not diminish the dataset’s benchmark status, but they shape its interpretation. The limited number of densely labeled anomaly masks suggests that methodological claims should be considered in light of annotation cost and sample efficiency. The absence of formal inter-annotator-agreement reporting means that label uncertainty must be inferred indirectly from the consensus workflow rather than quantified directly. The held-out split design emphasizes fixed-benchmark comparability over resampling-based variance estimation. The lack of described augmentation suggests that reported robustness is intended to arise primarily from model design and from the intrinsic diversity of Clinical and Wild acquisitions rather than from synthetic perturbation pipelines (Yu et al., 3 Jun 2026).

Taken together, ML-SASD occupies a specific niche in ocular image analysis: a benchmark centered on multi-label scleral anomaly segmentation under realistic acquisition variability, with dense pixel-wise supervision for three anomaly categories and explicit treatment of specular reflection as a structured source of false positives.