The application of deep learning to scalp analysis has gained significant traction. Early works focused on general scalp problem classification using CNN-based architectures [8,9] or robust representation learning in edge-cloud systems [10]. More recent work has shifted focus towards the more nuanced task of grading the severity of these conditions. Some approaches have repurposed object detection models, inferring severity from the density of detected problem areas [11,12]. However, this indirect method can be biased by non-scalp features or fail to capture fine-grained, fragmented symptoms. Consequently, a more direct approach based on established medical grading standards [13] has become prevalent. Jhong et al. enhanced a CNN with attention mechanisms [2], while Jin et al. [3] fine-tuned an EfficientNet for fine dandruff classification. Other works have explored ensemble models for robustness on limited data [5]. Or employed Vision Transformers (ViT) for grading multiple scalp conditions simultaneously [4]. While these studies demonstrate the potential of deep learning, they predominantly treat severity grading as a standard classification task. This overlooks two critical challenges: (1) the labels are ordinal in nature, and standard classification losses fail to capture this intrinsic order; (2) the labels are inherently noisy due to subjective expert assessment, which can degrade model performance. Our work explicitly addresses these two challenges in tandem.

Ordinal classification (or regression) aims to solve tasks where labels possess a natural ranking. A prominent and modular strategy involves decomposing a K-class ordinal problem into K-1 independent binary classification sub-problems [14]. This rank-based approach is particularly adept at addressing the challenge of ambiguous instances, which are samples that fall near the blurred boundaries between adjacent categories and are a common issue in subjective grading. Foundational deep learning models like MO-CNN [14] first operationalized this concept by training multiple binary classifiers to distinguish adjacent ranks. More recent methods, such as CORAL [15], have refined this approach to enforce rank consistency with greater efficiency. For contrast, other families of methods, such as Distribution Ordering Learning, seek to model the ordinal relationship by learning the label distribution directly. Techniques in this category include using soft labels [16]. While effective in certain contexts, these distribution-based methods often rely heavily on the precise location of the ground-truth label to anchor the distribution. This can make them potentially more sensitive to the high degree of label noise present in our dataset, where the ground-truth label itself is often unreliable. Therefore, for our framework, we adopt the rank-based decomposition strategy. Its proven effectiveness, modularity, and high compatibility with other learning paradigms make it an ideal backbone for integrating the noisy label learning methods we leverage.

To combat the memorization effect of deep networks on incorrect labels, various Learning with Noisy Labels (LNL) strategies have been proposed. These approaches can be broadly categorized into methods like Loss Adjustment and Sample Selection. Loss adjustment methods, such as Bootstrapping [17], attempt to correct the training signal for all samples, for example by using the model’s own predictions as a soft target. However, these correction-based approaches risk error propagation, especially in high-noise environments; if the model’s initial predictions are wrong, the corrected label will also be wrong, potentially reinforcing the error. In contrast, sample selection methods are particularly effective and relevant to our work. They operate on the principle that models learn from clean examples with small losses before fitting to noisy examples with large losses. Early methods in this domain, such as Co-teaching [18], utilized multiple networks to cross-filter clean samples for each other, mitigating confirmation bias. A leading approach, DivideMix [19], operationalizes this by modeling the per-sample loss distribution with a Gaussian Mixture Model (GMM) to dynamically separate the dataset into a likely clean set and a noisy set. Instead of discarding the noisy data, DivideMix reframes the task as a semi-supervised problem, using model predictions as pseudo-labels for the noisy subset. To further enhance feature learning on this subset without relying on potentially flawed pseudo-labels, contrastive learning has proven powerful for learning robust representations from the underlying images alone [20].

While these fields have advanced independently, a framework that synergistically integrates them is non-trivial and remains a critical gap. Ordinal regression methods typically assume that training labels are accurate [15,16]. Conversely, state-of-the-art LNL frameworks are class-agnostic [19] and do not enforce ordinal constraints. For instance, a standard class-agnostic loss metric, as used in many LNL methods, cannot differentiate minor from major ordinal errors, which is a vital distinction for robustly partitioning the data. Furthermore, reliance on pseudo-labeling for noisy data, a common LNL strategy, risks propagating ordinally-inconsistent errors. Our work bridges this gap by proposing the first unified framework built upon a rank-based ordinal backbone. This framework introduces two key innovations: it uses an ordinally-aware metric for robust sample partitioning and employs a dual-objective with contrastive learning to mitigate pseudo-label error propagation.

The task of grading dandruff severity presents dual challenges: the ordinal nature of severity levels and the prevalence of label noise. To address these, we propose a robust three-phase training framework that synergizes ordinal regression with a cooperative noisy label learning strategy. Our pipeline, illustrated in Fig. 1, progressively refines the model’s capability to handle noisy ordinal data. The process begins with a crucial supervised pre-training phase. Two parallel networks are independently warmed-up on the entire noisy dataset using an ordinal regression loss. This crucial first step ensures the models develop an initial understanding of the ranking relationships between severity grades, which is a prerequisite for generating meaningful loss distributions for sample partitioning. Leveraging these pre-trained models, the framework then proceeds to a cooperative sample partitioning phase. At the start of each subsequent epoch, the two networks use each other’s per-sample loss distributions to dynamically divide the dataset into a high-confidence clean subset and a low-confidence noisy subset. This cross-supervision or co-training design is critical for mitigating the confirmation bias inherent to self-training systems. With the data partitioned, the framework enters its main hybrid training phase, which employs a dual-objective learning strategy tailored to these subsets. A semi-supervised ordinal loss is applied to data generated via MixUp from both subsets, allowing the model to learn from the entire dataset. Concurrently, a label-agnostic contrastive loss is applied exclusively to the noisy subset to learn robust, generalizable features without relying on their corrupt labels. The subsequent sections provide a detailed exposition of each component. The complete procedure is formalized in Algorithm 1.

Standard multi-class classification, which typically employs a softmax output with categorical cross-entropy loss, is suboptimal for severity grading as it treats labels as independent nominal entities. This approach ignores the critical ordinal relationship between grades (e.g., grade 4 is more severe than grade 2). To formally integrate this structure, we adopt a rank-based methodology inspired by [15]. We reframe the K-class ordinal problem into K−1 simpler binary classification tasks. Specifically, a ground-truth label y∈{0,1,…,K−1} is transformed into a binary vector y′∈{0,1}K−1, where the k-th element indicates whether the severity grade is greater than rank k: (1)y′(k)=I(y≥k)fork∈{1,…,K−1}

To guarantee rank consistency in predictions (e.g., P(y≥4)>P(y≥2)), we constrain the network such that the final K−1 output neurons share the weights of the penultimate feature layer but each maintains an independent bias term. The model is then trained by minimizing the sum of binary cross-entropy (BCE) losses across all tasks. For a batch of B samples, this supervised loss Ls, is defined as: (2)Ls=1B∑i=1B∑k=1K−1BCE(σ(g(xi,θ)+bk,yi′(k))where g(xi,θ) denotes the feature vector from the penultimate layer, and bk is the bias for the k-th task. Minimizing this objective ensures that the learned biases are monotonically non-increasing (b1≥b2≥⋯≥bK−1), thereby guaranteeing rank-consistent probability estimates. This ordinal loss serves as the core objective for the pre-training (Phase 1) and the supervised component of the hybrid training (Phase 3).

A significant challenge in dandruff severity grading is the subtle visual distinction between adjacent levels, leading to inevitable label noise that can degrade model generalization. To address this, we employ a robust training strategy inspired by semi-supervised learning [19]. Instead of directly correcting labels, our approach dynamically partitions the data into a trusted clean set and an untrusted noisy set.

We posit that for a well-trained network, correctly labeled samples tend to exhibit lower losses than mislabeled ones. To formalize this, at the beginning of each epoch, we compute the per-sample ordinal loss for all training data. We then fit a two-component Gaussian Mixture Model (GMM) to the loss distribution to statistically separate clean and noisy samples. This process yields a posterior probability wi=P(clean|li) for each sample xi. To prevent error accumulation from a single model (self-confirmation bias), we implement a cooperative “co-training” framework. Specifically, the partition derived from Model 1’s loss distribution is used to train Model 2, and vice-versa. Samples with wi exceeding a threshold τ form the clean set Xc while the remainder form the noisy set Xn for which the original labels are discarded.

For samples in Xc, we further mitigate potential noise by applying a label refinement mechanism, as shown in Fig. 2. We compute a “soft” target y¯i′ by linearly combining the original ground-truth label yi′ with the model’s own prediction pi′ (averaged over augmentations):(3)y¯i′=wi⋅yi′+(1−wi)⋅pi′

This mechanism adaptively balances supervision: it trusts the provided label for high-confidence samples (where wi≈1) while relying more on the model’s prediction for ambiguous cases.

With the data partitioned into a clean set Xc (using refined labels y¯i′) and an unlabeled noisy set Xn, we employ a hybrid training strategy to maximize feature learning. This strategy combines a semi-supervised ordinal objective with a label-agnostic contrastive objective. First, we generate targets for the noisy subset Xn. As illustrated in Fig. 3, we compute a robust ensemble prediction by averaging the outputs of both co-trained networks, serving as the pseudo-label for samples in Xn. We then apply MixUp augmentation [21] to construct interpolated training batches combining refined samples from Xc and pseudo-labeled samples from Xn. The supervised ordinal loss Ls is minimized on these mixed batches.

To prevent the model from collapsing to a trivial solution (e.g., predicting a single class for all inputs), we incorporate a regularization term Lr. This forces the moving average of the model’s predictions pmodel¯ to align with the prior class distribution of the dataset pgt: (4)Lr=KL(pgt∣∣pmodel¯)

Concurrently, to extract discriminative features from the noisy data without relying on potentially erroneous pseudo-labels, we apply a contrastive loss Lu exclusively toXn. We utilize the InfoNCE objective to maximize the similarity between two augmented views (zi,zi+) of the same image while minimizing similarity with other samples:(5)Lu=−log⁡esim(zi,zi+)K∑j=1NnIj≠iesim(zi,zb)Kwhere sim(⋅,⋅) denotes cosine similarity, K is the temperature hyperparameter, and Nn is the batch size of the noisy subset. The total training objective is a weighted summation:(6)Ltotal=Ls+λu⋅Lu+λr⋅Lrwhere λu and λr are balancing hyperparameters.

To facilitate this research, we constructed a new, large-scale dataset for dandruff severity grading. The data was collected in collaboration with Taipei Hospital, Ministry of Health and Welfare and MacroHI Co., Ltd. in Taiwan, ensuring a diversity of clinical settings. This study was conducted in strict accordance with the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of Taipei Hospital, Ministry of Health and Welfare (Protocol No. TH-IRB-0022-0021). Using a digital microscope, clinicians and therapists captured images from eight distinct scalp regions, resulting in varied resolutions (768 × 576, 800 × 600, and 1024 × 768). A team of one dermatologist and three trained scalp therapists annotated a total of 22,537 images. To ensure a consistent and high-quality standard across this large-scale dataset, a rigorous adjudication protocol was employed. The trained therapists performed the initial annotations based on the clinically validated Adherent Scalp Flaking Score (ASFS) standard [13]. The dermatologist then served as the final expert adjudicator, reviewing annotations and defining the ground-truth label for each image used in the study. Each image was assigned to one of six severity levels {0, 2, 4, 6, 8, 10} based on the ASFS standard. The dataset exhibits a natural class imbalance representative of a clinical population: grade 0 (5817 images), grade 2 (9062), grade 4 (2966), grade 6 (2085), grade 8 (1893), and grade 10 (714). We randomly partitioned the data, allocating 18,871 images (~85%) for training and 3666 for testing, ensuring a consistent class distribution across splits. Fig. 4 shows example images for different severity grades. The creation of this multi-site, expert-annotated dataset is a key contribution to our work, providing a challenging and realistic benchmark for this task. Following standard practice, we report overall Accuracy. To provide a more robust assessment of our imbalanced dataset, we also report macro-averaged Precision, Recall, and F1-Score. The F1-Score is particularly important as it provides a balanced measure of performance across all six severity classes.

All experiments were conducted on an NVIDIA GeForce RTX 4090 GPU using a ResNet-50 backbone. For our method, the backbone was modified in two ways: the final classification layer was changed to have 5 outputs to suit the ordinal regression task, and a 128-dimensional projection head was added in parallel to facilitate contrastive learning. For data preprocessing, all input images were resized and randomly cropped to 224 × 224. Across all experiments, we utilized the Adabelief optimizer with β1=0.9 and β2=0.999, paired with a cosine annealing learning rate schedule. The training protocols were set as follows. For baseline and SOTA comparisons, models were trained for 450 epochs with a batch size of 64 and an initial learning rate of 1×10−4. Our proposed method involved a 10-epoch warm-up phase, followed by 250 epochs of hybrid training with a batch size of 32, an initial learning rate of 1×10−3, and a weight decay of 1×10−5. The key hyperparameters for our framework were set as: GMM partition threshold τ = 0.5, contrastive learning temperature K = 0.05, and regularization loss weight λr = 0.01. The contrastive loss weight λr was linearly ramped up from 0 to 0.06 over the first 100 epochs to stabilize initial training.

We compare our framework against several SOTA methods, as shown in Table 1. As the literature on robust ordinal learning for dandruff grading is nascent, we benchmark against high-performing models from related scalp analysis tasks [2–5,8]. All competing methods were trained on our dataset using the parameter settings reported in their original papers. As shown in Table 1, our proposed method significantly outperforms all other approaches across all metrics. Our ResNet-50-based model achieves an Accuracy of 80.71% and an F1-Score of 76.86%. This represents a substantial improvement of 2.26% in accuracy and 1.25% in F1-score over the next best method, the Triple ensemble model proposed by Kim et al. [5]. Notably, our single, principled framework surpasses not only this complex ensemble approach but also advanced architectures such as the Vision Transformer (ViT-B/16). These results validate that our framework’s core principles of explicitly modeling ordinal relationships and robustly handling label noise provide a decisive advantage over methods relying solely on architectural sophistication or ensembling.

To verify the contribution of each component of our framework, we conducted a detailed ablation study. Starting with a standard supervised ResNet-50 baseline, we incrementally integrated Ordinal Regression (OR), our Self-Supervised noisy label learning strategy (SSL), and Contrastive Learning (CL). The quantitative results are summarized in Table 2. The baseline model achieves a respectable accuracy of 78.23%. Introducing the OR formulation provides immediate gains, confirming the importance of modeling the inherent ordinal structure of severity grades. Subsequently, incorporating the SSL strategy with cooperative partitioning yields the most significant performance leap. As detailed in the per-class F1-scores, this gain is largely driven by a substantial improvement in Class 0 (rising from 75.68% to 80.03%), strongly suggesting that the “no dandruff” category contains considerable label noise which our SSL mechanism successfully mitigates. Finally, the addition of the CL module provides a further performance boost, achieving the highest Accuracy (80.71%) and Macro F1-score (76.86%). To rigorously assess performance on our imbalanced dataset, we also evaluated the Weighted F1-score. The proposed method achieves a Weighted F1 of 80.56%, surpassing the baseline’s 78.18%. This metric confirms that our framework maintains robust classification capabilities across all severity levels, rather than bias towards majority classes.

To further investigate the sources of these performance gains, we visualized the confusion matrices for the key ablation stages in Fig. 5. As observed in Fig. 5a, the Baseline model exhibits a dispersed prediction pattern, with misclassifications frequently spanning across non-adjacent classes (e.g., confus-ing Grade 2 with Grade 6). Incorporating the OR module (Fig. 5b) effectively constrains these predic-tions, resulting in a concentration of values along the diagonal and a significant reduction in large error margins. Furthermore, Fig. 5c,d illustrates the distinct contributions of the self-supervised strategies. The standalone SSL strategy (Fig. 5c) significantly corrects misclassifications in the noisy Grade 0 class. Subsequently, the intermediate OR + SSL stage (Fig. 5d) demonstrates how combining these approaches begins to align noise robustness with rank consistency. Finally, Fig. 5e presents the full Proposed Method (OR + SSL + CL). By integrating the contrastive learning objective, the model achieves the clearest diagonal structure with minimal off-axis errors. This visual evidence corroborates that our cooperative hybrid framework synergistically mitigates label noise while preserving the intrinsic ordinal structure.

Beyond classification performance, we also addressed the practical requirements for clinical deployment by evaluating inference speed (Runtime) as reported in Table 2. While our hybrid framework introduces a marginal increase in computational cost compared to the baseline (0.00476 s vs. 0.00264 s per image) due to the architectural design, the system remains highly efficient. With an inference speed capable of processing over 200 frames per second, our method proves to be computationally lightweight and well-suited for real-time diagnostic applications.