Patch Progression Masked Autoencoder (PPMAE)

• The paper introduces PPMAE, a masked autoencoder that predicts future OCT patches instead of reconstructing the same image, emphasizing temporal disease progression.
• It employs a ViT encoder with a fixed 75% random masking strategy and an asymmetric decoder to generate future-time patches for subsequent AMD evolution classification.
• Integrated into a two-stage pipeline, PPMAE uses generated future slices alongside a fusion CNN classifier to improve performance in classifying AMD progression stages.

Searching arXiv for the named paper and closely related masked-autoencoder masking-strategy work. Patch Progression Masked Autoencoder (PPMAE) is a masked-autoencoding framework introduced for forecasting structural change in optical coherence tomography (OCT) within the Monitoring Age-related Macular Degeneration Progression in Optical Coherence Tomography (MARIO) challenge. In its reported formulation, PPMAE does not reconstruct the same-time input image; instead, it predicts future-time patches from a follow-up OCT slice and then passes the pair consisting of the baseline slice and the generated future slice to a separate evolution classifier. The resulting system is a two-stage pipeline for exudative age-related macular degeneration (AMD): future OCT generation followed by progression classification into one of four labels, namely Reduced, Stable, Worsened, or Other (Zhang et al., 27 Aug 2025).

1. Clinical task and formal problem statement

Age-related Macular Degeneration (AMD) is described as a prevalent eye condition affecting visual acuity, and anti-vascular endothelial growth factor (anti-VEGF) treatments are reported as effective in slowing the progression of neovascular AMD, with better outcomes achieved through timely diagnosis and consistent monitoring. The MARIO challenge operationalized this monitoring problem through two prediction tasks on OCT data (Zhang et al., 27 Aug 2025).

In Task 1, the system receives two 2D OCT slices acquired at times $t$ and $t+1$, denoted $X_t$ and $X_{t+1}$, and predicts one of four evolution labels

$$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$

The classifier $f_1$ outputs probabilities

$p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$

In Task 2, only the baseline scan $X_t$ is given. The system first predicts the OCT that would be acquired three months later,

$\widehat Y = G(X_t),$

and then reuses the Task 1 classifier to categorize the evolution between $X_t$ and $t+1$0. The notation used in the reported method fixes $t+1$1, with preprocessing to $t+1$2 pixels for the classification pipeline, while $t+1$3 serves as the true follow-up slice or reconstruction target for PPMAE. The mask variable is written as $t+1$4, denoting the indices of masked patches (Zhang et al., 27 Aug 2025).

This formulation makes PPMAE a temporal masked autoencoder rather than a conventional masked image model. A plausible implication is that its pretext task is aligned with disease progression rather than generic image completion, because the target patches come from a later clinical acquisition rather than from the original input slice.

2. PPMAE architecture and reconstruction objective

PPMAE adapts the masked autoencoder paradigm to predict future-time patches rather than reconstructing the same-time image. The input $t+1$5 is resized to $t+1$6 and split into non-overlapping $t+1$7 patches. No overlap is used. This yields

$t+1$8

patches (Zhang et al., 27 Aug 2025).

The masking strategy is fixed and stochastic rather than learned. Specifically, the method randomly masks $t+1$9 of the patches, using a mask ratio

$X_t$0

and the set of masked patch indices $X_t$1 is sampled uniformly at random among all subsets of size $X_t$2. This point is important because it distinguishes PPMAE from adaptive masking methods that learn or schedule the mask policy (Zhang et al., 27 Aug 2025).

The encoder-decoder design is asymmetric. The encoder is a ViT-Large with 24 transformer blocks, embedding dimension $X_t$3, MLP dimension $X_t$4, and 16 attention heads. Only the unmasked patches are retained, augmented with standard learnable position embeddings, and processed by the encoder. The decoder has 8 transformer blocks, model dimension $X_t$5, and 16 heads. Its input contains the encoded visible-patch tokens together with learned mask tokens inserted at the masked positions, using the same positional embeddings. A final linear head projects decoder outputs back to patch-pixel predictions (Zhang et al., 27 Aug 2025).

The reconstruction target is the future slice $X_t$6 rather than the present slice $X_t$7. If the true future patches from $X_t$8 are denoted $X_t$9 and the decoder predictions are $X_{t+1}$0, the loss is the mean-squared error over masked patches:

$X_{t+1}$1

No extra progression loss or regularizer was used (Zhang et al., 27 Aug 2025).

Within the reported system, this design places all temporal supervision into the masked reconstruction target itself. This suggests that the model is encouraged to encode longitudinal retinal change through patch prediction alone, with no auxiliary loss shaping the latent space.

3. Fusion CNN classifier and the two-stage inference pipeline

The second stage of the overall system is a fusion CNN network for Task 1 evolution classification, built on a ResNet50 backbone pretrained on ImageNet. Two fusion strategies were evaluated: Early Fusion and Late Fusion (Zhang et al., 27 Aug 2025).

In Early Fusion, $X_{t+1}$2 and $X_{t+1}$3 are concatenated along the channel axis, producing an input of shape $X_{t+1}$4. This tensor is passed through ResNet50 to obtain a feature vector of size 2048, followed by a fully connected classification head producing four logits (Zhang et al., 27 Aug 2025).

In Late Fusion, each grayscale slice is replicated to 3 channels and processed separately by two identical ResNet50 encoders. This yields two 2048-dimensional vectors, which are concatenated into a 4096-dimensional fusion vector and passed through a single fully connected layer to obtain the four logits (Zhang et al., 27 Aug 2025).

The classification loss is cross-entropy. With $X_{t+1}$5 denoting the 4-dimensional logit output and $X_{t+1}$6 the one-hot label, the loss is

$X_{t+1}$7

For Task 2, the integration step is direct: once PPMAE has produced $X_{t+1}$8, the pair $X_{t+1}$9 is treated exactly as a Task 1 input, and the probabilities are computed as $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$0. The classifier $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$1 was trained using 4-fold cross-validation on the Task 1 training set, producing four models whose softmax outputs were averaged at inference time. No ensembling was applied to the PPMAE generator itself (Zhang et al., 27 Aug 2025).

A common misconception is to treat PPMAE as a standalone progression classifier. In the reported system, it is not. PPMAE generates the future slice, but the final progression label is produced by the separate Task 1 fusion classifier (Zhang et al., 27 Aug 2025).

4. Data processing, augmentation, and optimization

The preprocessing pipeline applied OCTIP flattening and retinal ROI extraction via inner limiting membrane alignment to all inputs. This preprocessing is reported across both tasks and is associated with improved validation metrics in the Task 1 classifiers and lower reconstruction error in PPMAE (Zhang et al., 27 Aug 2025).

Task 1 augmentations were RandomHorizontalFlip, RandomVerticalFlip, RandomRotation, ColorJitter, RandomPerspective, and GaussianBlur. PPMAE augmentations were RandomResizedCrop $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$2 and RandomHorizontalFlip (Zhang et al., 27 Aug 2025).

Optimization used AdamW with no scheduler. For Task 1 and Task 2 finetuning, the learning rate was $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$3, except for Task 2 finetuning where it was $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$4; batch size was 128; training ran for 150 epochs; and finetuning Task 1 on Task 2 data used 1 epoch. PPMAE itself was trained with learning rate $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$5, batch size 128, and 100 epochs (Zhang et al., 27 Aug 2025).

These settings emphasize that the reported gains are not tied to elaborate optimizer schedules or auxiliary losses. A plausible implication is that the central design choice is the temporal reconstruction target, not a heavily engineered training recipe.

5. Reported empirical results

For Task 1 classification on the validation set, OCTIP preprocessing consistently improved all metrics. Early Fusion without OCTIP yielded $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$6, $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$7, and Specificity $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$8, while Early Fusion with OCTIP yielded $$y \in \{\mathrm{Reduced}, \mathrm{Stable}, \mathrm{Worsened}, \mathrm{Other}\}.$$9, $f_1$0, and Specificity $f_1$1. Late Fusion without OCTIP produced $f_1$2, $f_1$3, and Specificity $f_1$4, whereas Late Fusion with OCTIP achieved the best reported Task 1 values, with $f_1$5, $f_1$6, and Specificity $f_1$7 (Zhang et al., 27 Aug 2025).

For PPMAE reconstruction on the Task 1 validation set, the reported mean-squared error was $f_1$8 for MAE without OCTIP, $f_1$9 for PPMAE without OCTIP, and $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$0 for PPMAE with OCTIP. The reported interpretation is that PPMAE outperforms standard MAE by a large margin and that OCTIP further halves the MSE (Zhang et al., 27 Aug 2025).

For Task 2 progression classification, the baseline composed of MAE plus Task 1 models yielded $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$1, $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$2, Specificity $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$3, $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$4, and Mean $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$5. Replacing MAE with PPMAE improved these values to $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$6, $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$7, Specificity $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$8, $p = \mathrm{softmax}(f_1(X_t, X_{t+1})).$9, and Mean $X_t$0. The strongest reported Task 2 result came from PPMAE plus Task 1 models finetuned on Task 2 data, with $X_t$1, $X_t$2, Specificity $X_t$3, $X_t$4, and Mean $X_t$5 (Zhang et al., 27 Aug 2025).

The team reported placement in the Top 10 for both challenge tasks. The same source also states that some team members were part of the same organization as the challenge organizers and were therefore not eligible to compete for the prize (Zhang et al., 27 Aug 2025).

6. Relation to adaptive masking and “patch progression” in adjacent MAE research

PPMAE uses uniformly random masking with a fixed ratio of $X_t$6, so its notion of progression resides in temporal prediction of future OCT patches rather than in the evolution of the mask policy itself (Zhang et al., 27 Aug 2025). In adjacent MAE literature, however, patch progression often refers to progressively changing which patches are masked or how masking difficulty evolves during training.

AutoMAE learns where to mask through a fully differentiable framework in which a Gumbel-Softmax mask generator is adversarially trained and interlinked with MAE reconstruction. It masks the top $X_t$7 of patches ranked by learned mask weights, uses a small 3-layer CNN discriminator, and reports on ImageNet-1K with ViT-B/p16 that linear probing improves from $X_t$8 to $X_t$9, while downstream gains on COCO and ADE20K are modest but positive (Chen et al., 2023).

CL-MAE introduces an explicit easy-to-hard curriculum via a learnable masking module. The module transitions smoothly from a partner of MAE, optimizing the same reconstruction loss, to an adversary, optimizing the opposite loss, under a curriculum factor $\widehat Y = G(X_t),$0 that decays from $\widehat Y = G(X_t),$1 to $\widehat Y = G(X_t),$2. On ImageNet, the reported nearest-neighbor accuracy for ViT-B rises from $\widehat Y = G(X_t),$3 for MAE to $\widehat Y = G(X_t),$4 for CL-MAE, and the method is reported to improve transfer performance across five downstream tasks (Madan et al., 2023).

Self-Guided MAE (SG-MAE) begins from an analysis of standard MAE and reports that patch embeddings form visually meaningful clusters early in training. It then replaces naive random masking with internally generated informed masks derived from normalized-cut bi-partitioning of encoder features, switching at an epoch $\widehat Y = G(X_t),$5 when decoder information flow from mask tokens surpasses that from visible tokens. Reported results include linear-probe gains from $\widehat Y = G(X_t),$6 to $\widehat Y = G(X_t),$7 after 400 epochs and training overhead of at most approximately $\widehat Y = G(X_t),$8 (Shin et al., 26 Jul 2025).

These methods are related but not equivalent. AutoMAE, CL-MAE, and SG-MAE modify mask selection during pretraining; PPMAE keeps the mask random and instead changes the reconstruction target from present-time content to future-time retinal content. A plausible implication is that “patch progression” can denote either temporal progression in the prediction target, as in PPMAE, or progressive adaptation of the masking policy, as in these related MAE variants.

7. Conceptual significance and technical boundaries

Within the reported MARIO solution, PPMAE is significant because it converts a future-progression problem into a generative intermediate task followed by a discriminative classification task. The future slice $\widehat Y = G(X_t),$9 serves as a surrogate follow-up, allowing a Task 1 evolution classifier to be reused for Task 2 without changing its output space or loss function (Zhang et al., 27 Aug 2025).

At the same time, the technical boundaries of the reported method are explicit. PPMAE uses random masking rather than informed masking, uses only an MSE reconstruction loss over masked patches, applies no extra progression loss or regularizer, and does not ensemble the generator. The strongest Task 2 classification result is obtained only after finetuning the Task 1 classifiers on Task 2 data, which indicates that the discriminative stage remains materially important even when the generative stage is improved (Zhang et al., 27 Aug 2025).

This evidence supports a narrow but precise characterization: PPMAE is not primarily a new masking-policy learner, nor a generic OCT synthesizer in isolation, but a task-specific temporal masked autoencoder embedded in a two-stage clinical progression pipeline.