Dynamic Curriculum Learning for Spatiotemporal Encoding

• The paper demonstrates that DCL-SE leverages a two-stage approach—compact 3D-to-2D encoding via Approximate Rank Pooling followed by dynamic curriculum decoding—to achieve state-of-the-art performance with fewer parameters.
• The method utilizes a spatiotemporal encoding mechanism that preserves anatomical progression and fine-grained pathological details, outperforming conventional projections and resource-intensive 3D CNNs.
• The curriculum decoding with adaptive grouped convolutions mimics clinical reasoning by progressively refining features, which significantly enhances multi-task outcomes like classification, segmentation, and regression.

Dynamic @@@@2@@@@ for Spatiotemporal Encoding (DCL-SE) is an end-to-end neural framework designed for high-dimensional neuroimaging analysis. It addresses the limitations of conventional clinical pipelines that either project 3D volumetric data to individual 2D slices, losing anatomical and developmental cues, or rely on resource-intensive 3D CNNs that are computationally demanding and require large annotated datasets. DCL-SE leverages a two-stage paradigm: compactly encoding volumetric progression into a single dynamic image via Approximate Rank Pooling (ARP), then decoding through a dynamic curriculum guided by spatiotemporal complexity metrics and adaptive grouped convolutions. This data-driven approach progressively refines features from global anatomical structures to fine-grained pathological details and achieves state-of-the-art performance across diverse tasks, such as disease classification, segmentation, and regression, with markedly fewer parameters than large generic models (Zhou et al., 19 Nov 2025).

1. Problem Formulation and Foundations

DCL-SE is motivated by the constraints in neuroimaging, specifically the compromise between spatiotemporal fidelity and tractable model design for clinical diagnostics. The conventional practice of projecting 3D MRI/CT volumes onto 2D slices breaks anatomical continuity and fails to exploit sequential progression signals crucial for clinical phenotyping. Pure 3D convolutional models preserve spatial context but are hampered by computational cost and annotation scarcity.

Static curriculum learning, structured as hand-engineered sequences from "easy" to "hard" tasks, is inadequate for neuroimaging, where the complexity of anatomical features evolves dynamically across scales and tasks. DCL-SE addresses this gap through:

• A spatiotemporal encoding mechanism that compresses 3D volumes into information-rich 2D dynamic images, maintaining progression cues.
• A decoder trained via a curriculum that dynamically adapts to the scale and complexity of features extracted.

These advances emphasize that task-adaptive, compact architectures can outperform generic, large-scale pretrained models in accuracy, robustness, and interpretability (Zhou et al., 19 Nov 2025).

2. Spatiotemporal Encoding via Approximate Rank Pooling

The first stage of DCL-SE is Data-based Spatiotemporal Encoding (DaSE). Given a volume comprising $T$ ordered slices $\{I_1, \dots, I_T\}$, per-slice features $\psi_t = f_{\mathrm{enc}}(I_t) \in \mathbb{R}^d$ are extracted. Approximate Rank Pooling is then employed to aggregate these features into a single descriptor $d^*$ by solving a ranking SVM objective:

$$\min_d\;\frac{\lambda}{2}\|d\|^2 + \frac{2}{T(T-1)}\sum_{q>t}\max\{0,1 - S(q|d) + S(t|d)\}$$

where

$$S(t|d)=\langle d,V_t\rangle,\quad\text{with}\quad V_t=\frac{1}{t}\sum_{i=1}^t\psi_i$$

The closed-form solution yields: $$d^* = \sum_{t=1}^T \alpha_t \psi_t, \quad \alpha_t = 2(T-t+1) - (T+1)(H_T - H_{t-1})$$ where $H_t = \sum_{i=1}^t 1/i$. This $d^*$ is reshaped to form a single 2D dynamic image encapsulating the anatomical evolution across the volume (e.g., $H\times W$).

Replacing ARP with average or max pooling results in substantial performance degradation (e.g., F1 drops from 96.6% to 88–90%) (Zhou et al., 19 Nov 2025).

3. Dynamic Curriculum Decoding and Complexity-Adaptive Progression

The second stage, Curriculum Semantic-based Decoding, processes the dynamic image using a shallow 2D encoder and a decoder that operates in a sequence of curriculum stages ($S1\to S4\to S6\to C1$):

• Each stage produces feature maps $X_i\in\mathbb{R}^{C\times H\times W}$.
• A complexity metric guides progression:

$\lambda(X_i) = \sum_{c=1}^C \|\nabla X_i^{(c)}\|_1$

• Stage transition occurs when $\lambda(X_i)$ exceeds threshold $\tau_i$, empirically set by percentile statistics over the training set.

The curriculum mimics clinical reasoning by moving from coarse anatomical analysis (gray/white matter) to fine pathological boundary localization (small lesions). Removing the curriculum progression drops AUC by approximately 7 percentage points in classification experiments (Zhou et al., 19 Nov 2025).

4. Dynamic Group Mechanism (DGM) for Adaptive Feature Recalibration

At every stage, the Dynamic Group Mechanism adaptively recalibrates channel and spatial feature saliency through grouped convolutions:

1. Apply a channel-reduction convolution → ReLU → channel-expansion convolution → sigmoid to derive an importance map $W$.
2. Modulate feature maps: $X' = X \odot W$.
3. Further grouped pointwise convolutions fuse spatial detail, producing the output for the current stage.

Grouped convolution settings commonly use $G=4$ and reduction factor $r=16$. Disabling DGM at any stage degrades accuracy by 3–5 percentage points (Zhou et al., 19 Nov 2025).

5. Loss Functions, Training Protocols, and Task Heads

DCL-SE employs task-specific heads and loss functions:

• Classification: cross-entropy loss.
• Segmentation: composite loss

$$L=0.5\,L_{\mathrm{Lovasz\text{-}Soft}} + 0.3\,L_{\mathrm{Lovasz\text{-}Hinge}} + 0.2\,L_{\mathrm{Boundary}}$$

• Regression (e.g., brain age): L1 loss (Mean Absolute Error).

Optimization is performed using AdamW (weight decay from $1\times 10^{-2}$ to $1\times 10^{-5}$), with cosine-annealing or OneCycleLR learning rate schedules.

6. Evaluation, Benchmarks, and Ablations

DCL-SE is evaluated across six publicly available datasets encompassing Alzheimer's classification (AD-MRI, ADNI), brain tumor classification (BT-MRI/CT), cerebral vessel segmentation (CAS2023), and brain age prediction (SPR Head CT).

• Metrics include accuracy, AUC, F1, Dice Similarity Coefficient (DSC), MAE, parameter count, and FLOPS.
• DCL-SE achieves top or state-of-the-art results on all benchmarks, e.g.:
  • 3D Alzheimer's: AUC = 97.1%, F1 = 96.6%
  • AD-MRI: 99.94% accuracy with 8M parameters and 1.1 GFLOPS
  • Vessel segmentation: top mean DSC among 8 challenge teams
  • Age prediction: MAE ≈ 5.3 years, outperforming open-source models by more than 2 years
• Large foundation models (GPT-4, Gemini, Claude) achieve near-random performance (AUC ~ 0.55–0.65) on multi-class neuroimaging tasks without domain fine-tuning; DCL-SE approaches AUC ~ 0.98 with minimal adaptation.

7. Relation to General Spatiotemporal Curricula and Representation Learning

DCL-SE generalizes curriculum paradigms seen in spatiotemporal forecasting and video self-supervised learning. Temporal Progressive Growing Sampling (TPGS) (Liu et al., 2020) employs multi-scale curricula and dynamic mixing decay schedules for sequence forecasting, while Dense Predictive Coding (DPC) (Han et al., 2019) adopts curriculum stages by progressively reducing input context and lengthening future prediction horizons in video representation learning.

DCL-SE subsumes such approaches:

• By extending curriculum learning over spatial grids and temporal scales with adaptive mixing coefficients,
• By coupling stage transitions to data-driven complexity metrics rather than fixed schedules,
• By integrating grouped convolutional gating for explicit modulation of anatomical and pathological features.

These mechanisms enable DCL-SE to learn compact, order-preserving representations and robust feature hierarchies interpretable via t-SNE and GradCAM visualizations (Zhou et al., 19 Nov 2025).

Summary Table: DCL-SE Core Components

Component	Mathematical Formulation	Role in Pipeline
Approximate Rank Pooling (ARP)	$d^* = \sum_{t=1}^T\alpha_t\psi_t$	3D→2D spatiotemporal encoding
Curriculum Progression	$\lambda(X_i) = \sum_{c=1}^C \|\nabla X_i^{(c)}\|_1$	Dynamic stage transitions
Dynamic Group Mechanism (DGM)	$$W = \sigma(\mathrm{Conv}_G(\mathrm{ReLU}(\mathrm{Conv}_G(X))))$$	Adaptive feature recalibration

DCL-SE demonstrates that a carefully structured, progression-aware encoding and decoding pipeline can achieve superior performance and clinical utility in neuroimaging tasks compared to static architectures and generic large models (Zhou et al., 19 Nov 2025).