AGE-Net: Deep Learning in Medical Imaging

• AGE-Net is a suite of deep learning frameworks that combine multi-frequency spectral-spatial fusion with anatomical graph reasoning to analyze biomedical images.
• It incorporates a modular ConvNeXt-based pipeline for knee osteoarthritis grading and a hybrid 3D CNN for MRI-based brain age estimation.
• The architectures use evidential ordinal regression to quantify predictive uncertainty, achieving robust performance metrics in clinical assessments.

AGE-Net refers to multiple deep learning architectures proposed for biomedical image analysis, each tailored to a specific imaging modality and prediction task but united in their focus on learning aging-related or disease-related patterns from complex data. Prominent AGE-Net variants include: (1) a spectral–spatial fusion and anatomical-graph reasoning model for automated Kellgren–Lawrence (KL) grading on knee radiography (Li et al., 24 Jan 2026) and (2) a 3D CNN framework for brain MRI-based organ-specific biological age estimation with iterative cohort-cleaning (Armanious et al., 2020). Both exemplify recent advances in integrating spatial context, regularizing ordinal prediction, and quantifying model uncertainty for clinically relevant assessment.

1. AGE-Net Architectures: Overview and Task Formulation

AGE-Net for knee osteoarthritis (OA) grading employs a modular ConvNeXt-based workflow that integrates multi-frequency (spectral) and region-specific (spatial) representations, graph-theoretic long-range reasoning, and boundary-centric refinement. The model predicts ordinal KL grades (0–4) from 2D radiographs and incorporates evidential uncertainty through a Normal-Inverse-Gamma (NIG) head (Li et al., 24 Jan 2026).

In MRI-based biological age estimation, Age-Net uses a 3D hybrid architecture fusing Inception and SqueezeNet "fire" modules, optimized to regress chronological age (CA) from high-dimensional gray-matter segmentations. Iterative data cleaning identifies outlier subjects hypothesized to deviate from normal organ-specific biological age, thereby refining the mapping toward bona fide biological age (BA) inference (Armanious et al., 2020).

2. Spectral–Spatial Fusion and Anatomical Graph Reasoning for OA Grading

The AGE-Net for KL grading systematically augments a standard ConvNeXt-Base backbone with three enhancement modules:

• Spectral–Spatial Fusion (SSF): Decomposes backbone feature maps via real 2D Fourier transform (rFFT), modulates frequency bands with learnable complex weights, then reconstructs the fused tensor. Parallel spatial attention is computed via bottlenecked $1\times1$ convolutions and a sigmoid gate. SSF outputs:

$$\mathrm{SSF}(F) = F + F_{\mathrm{spec}} + (F \odot A)$$

where $F_{\mathrm{spec}}$ is the inverse-transformed spectral tensor and $A$ is the attention mask.

• Anatomical Graph Reasoner (AGR): Converts spatially pooled features into a fixed grid of anatomical tokens, builds a $k$-nearest neighbor graph in feature space ($k=9$), and passes messages using an EdgeConv-style MLP:

$$x'_i = \max_{j\in \mathcal{N}(i)} h_\theta\left([x_i, x_j - x_i]\right)$$

Outputs are projected back, upsampled, and gated for featurewise multiplication and residual addition.

• Differential Refiner (DFR): Applies depthwise $3\times3$ convolution, computes the absolute differential map $E = |F_d - F|$, mixes via $1\times1$ convolution and batch-norm, and fuses via a SiLU nonlinearity:

$\mathrm{DFR}(F) = \mathrm{SiLU}(R + F)$

These stages enable AGE-Net to exploit both global context and localized pathoanatomic markers such as joint space narrowing.

3. Evidential Ordinal Regression and Training Objectives

AGE-Net employs a NIG evidential regression head to model predictive mean $\gamma$ and uncertainty parameters $(\nu, \alpha, \beta)$:

• Output: $(\gamma, \nu, \alpha, \beta)$, with positive constraints (softplus + $\epsilon$).
• Predictive distribution: Student-$t$ parameterized by NIG, with variance

$$\mathrm{Var}(y) = \frac{\beta (1+\nu)}{\nu (\alpha-1)}, \quad \alpha>1$$

• Loss function:

$$\mathcal{L}_{\mathrm{evi}} = -\log p_{\mathrm{NIG}}(y|\gamma,\nu,\alpha,\beta) + \lambda(t) |y-\gamma| (2\nu+\alpha)$$

where $\lambda$ is linearly scheduled.

Ordinality is enforced by a pairwise margin ranking loss across valid label pairs $(i,j)$ with $y_i - y_j \geq 1$:

$$\mathcal{L}_{\mathrm{rank}} = \frac{1}{|\Omega|} \sum_{(i,j)\in\Omega} \max\left\{0,\, m - (\gamma_i - \gamma_j)\right\}$$

with margin $m=0.8$ and a weighting of $\alpha_\mathrm{rank}=2.0$ in the total loss.

4. AGE-Net for MRI-Based Biological Age Estimation

In the context of brain biological age prediction, Age-Net leverages a "hybrid" 3D CNN architecture:

• Building Blocks:
  • Four modules each comprising two 3D Inception blocks (multi-scale parallel convolutions) and a Fire module (SqueezeNet-style parameter efficiency).
  • All convolutions are followed by batch normalization, ReLU, and $L_2$ regularization.
  • Global average pooling and three fully connected layers, with subject gender concatenated before the final regression output.
• Input Modes:
  • Full-volume gray-matter segmentations ($121\times145\times121$ voxels).
  • Chunk-based mode, splitting into 14–20 slabs ($121\times145\times6$), per-chunk predictions averaged at inference.
• Iterative Outlier Removal:
  • On each round, patients whose mean chunk-prediction deviates from true CA beyond $R=1.96$ times the per-chunk standard deviation are flagged as outliers and removed.
  • After three stable iterations, the pruned cohort is presumed to satisfy BA $\approx$ CA; retraining yields a BA estimator.

5. Quantitative Performance and Ablation Results

Evaluation metrics and results demonstrate the effectiveness of AGE-Net’s architectural choices in both application domains:

Model Variant	QWK (KL Grading)	MSE (KL Grading)	MAE (Brain Age)	RMSE (Brain Age)	Corr. (Brain Age)
AGE-Net (full, knee OA)	0.9017 ± 0.0045	0.2349 ± 0.0028	—	—	—
w/o ranking (knee OA)	0.8963 ± 0.0033	0.2600 ± 0.0083	—	—	—
w/o AGR (knee OA)	0.8932 ± 0.0018	0.2665 ± 0.0017	—	—	—
Base ConvNeXt (knee OA)	0.8909 ± 0.0113	0.2795 ± 0.0287	—	—	—
Age-Net (full-vol, brain)	—	—	2.658	3.584	0.979
Age-Net (chunk+gender, brain)	—	—	1.955	3.210	0.983

In knee OA grading, each component (SSF, AGR, DFR, evidential head, ordinal ranking) is necessary for optimal QWK and MSE performance. In brain age estimation, chunk-based and gender-conditioned variants consistently deliver lowest MAEs relative to state-of-the-art MRI age regressors (Li et al., 24 Jan 2026, Armanious et al., 2020).

6. Uncertainty Quantification, Robustness, and Interpretability

AGE-Net’s evidential ordinal head enables direct estimation of predictive uncertainty, shown to correlate positively with absolute error. Risk–coverage analysis demonstrates enhanced performance in selective prediction settings. Perturbative robustness—evaluated under noise, blur, and contrast shifts—confirms that AGE-Net degrades more gracefully than conventional CNN baselines. AGR gates and DFR maps offer interpretable spatial focus on joint compartments and margins.

In the MRI setting, joint GradCAM++ and guided-backprop saliency elucidates network focus on canonical neuroanatomical regions (hippocampus, amygdala, thalamus). Alzheimer’s-labeled outliers manifest diminished activation in such regions; mild cognitive impairment cases display intermediate patterns, supporting explanatory value for saliency-informed BA shifts (Armanious et al., 2020).

7. Applications, Limitations, and Prospects

AGE-Net advances the state of the art in both automated disease grading (KL) and organ-specific biological age profiling from medical images. Clinical applications include differential OA grading for prognosis and therapy selection, as well as non-invasive BA estimation relevant to geriatrics, cognitive decline monitoring, and patient stratification.

Key limitations include: (a) reliance on proxy BA labels under the assumption that healthy subjects age “normally,” (b) potential exclusion of healthy but atypical individuals during cohort refinement, and (c) the challenge of domain adaptation beyond expert-annotated datasets. Proposed future directions include extension to whole-body MRI, integration of Bayesian uncertainty, longitudinal modeling, and exploration of unsupervised lesion/anomaly detection via learnable saliency (Li et al., 24 Jan 2026, Armanious et al., 2020).