PSScreen: Partially Supervised Retinal Screening

• PSScreen is a partially supervised framework for multi-label retinal disease screening that addresses missing labels and domain shifts across diverse medical datasets.
• The original model employs a two-stream architecture with deterministic and probabilistic branches, while PSScreen V2 extends this with a three-branch teacher-student system incorporating low-frequency dropout and uncertainty.
• Empirical evaluations show both versions achieve state-of-the-art performance and robust generalization across in-domain, unseen, and cross-modality datasets.

PSScreen is a partially supervised framework for multiple retinal disease screening from fundus photographs that was introduced to learn from multiple partially labeled datasets collected at different medical sites, where both label absence and domain shift are central obstacles. The original model uses a two-stream design with a deterministic stream and a probabilistic stream, textual guidance for disease-wise feature decoupling, feature distillation, pseudo-label consistency, and self-distillation; PSScreen V2 extends this line into a three-branch self-training framework with one teacher and two student networks, adding Low-Frequency Dropout and Low-Frequency Uncertainty to improve domain robustness under heterogeneous data distributions (Zheng et al., 14 Aug 2025, Zheng et al., 26 Oct 2025).

1. Problem formulation and disease space

PSScreen is defined for multi-label retinal screening over a meta-dataset assembled from multiple partially labeled datasets. In the original formulation, an image $x$ has a label vector $y \in \{-1,0,1\}^T$, where $y_t=0$ denotes unknown and $y_t=1/-1$ denotes positive/negative, with a binary mask $\delta_t = 1$ if $y_t \in \{-1,1\}$ and $0$ otherwise. PSScreen V2 uses a closely related but version-specific notation: $y \in \{1,0,-1\}^T$, where $y_t = 1/0$ denotes positive/negative and $y_t = -1$ denotes unknown, with an observed-label mask $y \in \{-1,0,1\}^T$0 (Zheng et al., 14 Aug 2025, Zheng et al., 26 Oct 2025).

The disease space is organized around seven categories: normal, diabetic retinopathy, glaucoma, cataract, age-related macular degeneration, hypertensive retinopathy, and pathologic myopia. The in-domain meta-dataset combines DDR, REFUGE2, PALM, ADAM, Kaggle-CAT, and Kaggle-HR, while out-of-domain evaluation includes APTOS2019, ORIGA$y \in \{-1,0,1\}^T$1, HPMI, RFMiD, and ODIR. PSScreen V2 further evaluates transfer on chest X-ray datasets, namely CheXpert, ChestX-ray14, and SIIM-ACR Pneumothorax2019.

The core difficulty is twofold. First, each constituent dataset provides labels only for a subset of diseases, so naïve fully supervised training is infeasible. Second, the datasets differ in camera, illumination, resolution, preprocessing, label definitions, and patient cohort, which induces substantial domain shift. Both versions of PSScreen treat these as coupled rather than separable problems.

2. Original PSScreen architecture

The original PSScreen is a two-stream model. One stream learns deterministic features and acts as the stable supervisory anchor; the other learns probabilistic features through Domain Shifts with Uncertainty blocks inserted after each backbone stage. In the probabilistic stream, instance statistics are perturbed by sampling $y \in \{-1,0,1\}^T$2 and $y \in \{-1,0,1\}^T$3, followed by instance normalization and re-styling,

$y \in \{-1,0,1\}^T$4

so that the student is exposed to plausible style-domain variations without requiring target-domain images (Zheng et al., 14 Aug 2025).

A second defining component is text-guided semantic decoupling. Disease descriptions derived from expert knowledge are encoded with BioClinicalBERT, producing disease embeddings $y \in \{-1,0,1\}^T$5. These are matched against the last-stage visual map through a disease-specific attention mechanism,

$y \in \{-1,0,1\}^T$6

and disease-wise features are obtained by attention-weighted pooling, $y \in \{-1,0,1\}^T$7. The same decoupling is applied to the probabilistic stream, yielding disease-aligned feature pairs for distillation.

Training combines four losses. The deterministic stream is supervised on known labels by $y \in \{-1,0,1\}^T$8. Disease-wise deterministic and probabilistic features are aligned by an MMD term $y \in \{-1,0,1\}^T$9. Unknown classes are handled by pseudo-label consistency $y_t=0$0, using deterministic predictions with confidence threshold $y_t=0$1. Known classes also induce a KL-based self-distillation term $y_t=0$2. The total objective is

$y_t=0$3

with $y_t=0$4, $y_t=0$5, $y_t=0$6, and $y_t=0$7 scheduled as $y_t=0$8 for the first 5 epochs and $y_t=0$9 from epoch 6 onward. At inference time, the default prediction is the simple average of the two streams’ probabilities.

3. PSScreen V2 and the frequency-domain redesign

PSScreen V2 replaces the two-stream layout with a three-branch architecture comprising one teacher and two students. The teacher branch consumes weakly augmented images and generates pseudo labels on the fly; it is explicitly not an EMA teacher. Student 1 applies Low-Frequency Dropout, and Student 2 applies Low-Frequency Uncertainty at each encoder block. The teacher provides pseudo labels for unknown classes, while Student 2 is additionally constrained by feature-level MMD and output-level KL consistency against the teacher (Zheng et al., 26 Oct 2025).

Low-Frequency Dropout operates in the Fourier domain on encoder feature maps. If $y_t=1/-1$0 with amplitude $y_t=1/-1$1 and phase $y_t=1/-1$2, a centered low-pass mask $y_t=1/-1$3 selects a square low-frequency region of side $y_t=1/-1$4, with $y_t=1/-1$5 in practice. A Bernoulli mask with drop probability $y_t=1/-1$6 is applied only inside the low-frequency band, and the modified amplitude is reconstructed by inverse FFT. The intended effect is to discard domain-related low-frequency components while preserving higher-frequency lesion cues.

Low-Frequency Uncertainty is more elaborate. At each encoder level, V2 computes channel-wise low-frequency means and variances, perturbs them by Gaussian noise with learnable scales $y_t=1/-1$7 and $y_t=1/-1$8, then re-styles the low-frequency amplitude by an AdaIN-like transformation: $y_t=1/-1$9 The noise scales are optimized adversarially through $\delta_t = 1$0, yielding the alternating optimization

$\delta_t = 1$1

The architectural contrast between the two versions is concise:

Version	Branch structure	Distinctive mechanisms
PSScreen	Deterministic stream + probabilistic stream	DSU, text-guided decoupling, MMD alignment, pseudo-label consistency, self-distillation
PSScreen V2	Teacher + Student 1 + Student 2	Low-Frequency Dropout, Low-Frequency Uncertainty, on-the-fly pseudo labels, MMD and KL consistency

For V2, pseudo labels are generated with $\delta_t = 1$2. Student 2 is trained by

$\delta_t = 1$3

with $\delta_t = 1$4, $\delta_t = 1$5, and $\delta_t = 1$6. Inference uses only the teacher branch.

4. Data processing, optimization, and implementation

The original PSScreen preprocesses each image by cropping the field of view, padding the short side, and resizing to $\delta_t = 1$7. Training uses random scaling in $\delta_t = 1$8 with probability $\delta_t = 1$9, plus augmentations from prior work excluding Cutout. Optimization uses Adam with batch size $y_t \in \{-1,1\}$0, weight decay $y_t \in \{-1,1\}$1, initial learning rate $y_t \in \{-1,1\}$2, decay by a factor of $y_t \in \{-1,1\}$3 every 10 epochs, for 20 epochs on one NVIDIA A100 (40GB) (Zheng et al., 14 Aug 2025).

PSScreen V2 retains the $y_t \in \{-1,1\}$4 preprocessing but adopts weak image augmentations for the teacher, random scaling uniformly in $y_t \in \{-1,1\}$5 with $y_t \in \{-1,1\}$6, and a Multi-label Transformer augmentation set excluding Cutout. Optimization changes to AdamW with batch size $y_t \in \{-1,1\}$7, weight decay $y_t \in \{-1,1\}$8, initial learning rate $y_t \in \{-1,1\}$9, decay by $0$0 every 10 epochs, for 20 epochs on two NVIDIA V100 32GB GPUs. To prevent small datasets from being marginalized, each batch draws equal numbers from each dataset in the meta set (Zheng et al., 26 Oct 2025).

The computational profiles are close to real-time screening scale rather than large foundation-model scale. For the original model, the reported efficiency is 373.1 FPS, 660 GFLOPs, and 116.8M parameters. For V2 with ResNet-101, the reported complexity is 661 GFLOPs, 177.9M parameters, and 371.6 FPS. This indicates that the move from DSU to frequency-domain perturbations increases model size more than raw throughput cost.

5. Empirical performance and generalization

Both versions are reported as state-of-the-art on their evaluation suites. The original PSScreen reaches 84.2% mF and 76.8% mQWK on the in-domain meta-dataset, 65.9% mF and 50.9% mQWK on unseen datasets, and 64.1% mF and 39.8% mQWK on ODIR. PSScreen V2 improves these to 84.5%/78.0%, 66.1%/51.6%, and 64.2%/39.9%, respectively (Zheng et al., 14 Aug 2025, Zheng et al., 26 Oct 2025).

Setting	PSScreen	PSScreen V2
Meta mF / mQWK	84.2 / 76.8	84.5 / 78.0
Unseen mF / mQWK	65.9 / 50.9	66.1 / 51.6
ODIR mF / mQWK	64.1 / 39.8	64.2 / 39.9

Several ablations clarify where the gains arise. In the original model, removing $0$1, $0$2, or $0$3 degrades performance, and using $0$4 alone can hurt because the student drifts without teacher guidance. In V2, the full teacher-plus-two-student system outperforms all reduced configurations; removing pseudo-label losses degrades performance; and removing $0$5 harms more than removing $0$6, which the paper attributes to the broader transfer of information through feature consistency. V2 also reports that $0$7 and $0$8 are optimal, while larger $0$9 damages high-frequency lesion cues.

The generalization claims extend beyond fundus images. On ODIR200×3, the original PSScreen outperforms FLAIR by 18.7% ACA, and V2/V1 outperform the best foundation model FLAIR by ≈20%/18.7% ACA. V2 also reports compatibility with ResNet-101, ConvNeXt V2-T, Swin-T, VMamba-T, and DINOv2 with PEFT methods such as SoMA, DoRA, FLoRA, and LoRA-Dash. On chest X-ray transfer, V2 achieves macro F-scores of 56.4% on CheXpert, 58.2% on ChestX-ray14, and 75.9% on SIIM-ACR Pneumothorax2019.

6. Interpretation, limitations, and research significance

The main conceptual claim shared by both versions is that partial supervision and domain generalization should be solved jointly. In the original model, this is achieved by pairing a deterministic teacher-like branch with a probabilistic branch perturbed by DSU and regularized through text-guided feature alignment. In V2, the same overall objective is reformulated in the frequency domain: low-frequency components are treated as carriers of domain style, whereas higher-frequency components are treated as more lesion-specific and disease-indicative. This suggests a shift from generic feature uncertainty modeling to a more explicit structural prior about where domain information resides (Zheng et al., 14 Aug 2025, Zheng et al., 26 Oct 2025).

The reported limitations are also version-specific but coherent. In the original system, hypertensive retinopathy remains difficult, and current experiments focus on 2D fundus photography rather than multimodal ophthalmic imaging. In V2, DR level-1 lesions are described as subtle and rare, pseudo labels can still be noisy when teacher confidence is miscalibrated, extreme domain shifts beyond the learned low-frequency perturbation family may degrade performance, and the teacher is not an EMA model. Neither version treats multimodal ophthalmic fusion as a solved problem.

Within retinal screening research, PSScreen is notable for making partially labeled multi-site training the primary design target rather than a side condition. Its progression from disease-wise text-guided distillation to adversarial low-frequency perturbation reflects a broader methodological trend: clinical screening models increasingly treat annotation sparsity, cross-site heterogeneity, and semantic priors as first-class components of the learning problem rather than post hoc nuisances.