- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes a spatially adaptive convolution in which the outputs of five heterogeneous 3D convolutional experts are combined with per-location softmax weights, predicted by a window-attention router in the encoder and a small convolutional router in the decoder. The formulation is embedded in an autoencoder trained to reconstruct T1-weighted MRI volumes from IXI. The paper focuses on characterizing expert behavior: routing maps, correlations between expert responses and image-structure measures, responses to sinusoidal inputs, and expert-removal ablations. It concludes that the experts show differentiated behavior consistent with functional specialization under a soft, spatially varying mixture, and reports higher reconstruction fidelity than a MONAI autoencoder and a 3D ViT.
Critical strengths
A good control for routing is included, and the analyses are qualified carefully.
- The paper compares adaptive routing against uniform aggregation of the same expert bank (Table 5, Appendix A.4), which is the correct control for isolating the effect of routing.
- It explicitly frames the structural and frequency analyses as correlational rather than causal, and reports that expert roles vary across random seeds (Appendix C.7) rather than presenting a single run as representative.
Critical weaknesses
1. There is no evidence that spatial routing improves the model, and the reconstruction comparisons are not interpretable.
- Routing adds little over uniform aggregation: The adaptive five-expert model reaches 29.04 dB PSNR and 0.9596 SSIM, against 28.84 dB and 0.9586 for uniform aggregation of the same experts (Table 5). No variance across seeds is reported for these comparisons, so there is no way to establish statistical significance.
- Simpler configurations match or exceed the full model: A single expert without routing (E2, a standard 1×3×3 convolution) reaches 29.14 dB and 0.9593 SSIM. Removing E0 or E3 from the full model improves PSNR to 29.68 and 29.64 dB.
- The learned routing is close to uniform: The normalized routing entropy of the full model is 0.957 (Table 2). In the encoder, the colorbars of Fig. 8 indicate that each expert's weight varies only within a narrow range across the slice (e.g., approximately 0.59–0.62 for expert 1). This amounts to an approximately fixed global weighting rather than meaningful spatial adaptation.
- yet the cost is substantial: Relative to the MONAI autoencoder, inference time is about 13× higher (347.79 vs. 26.35 ms) and peak memory about 7× higher (2.276 vs. 0.313 GB) (Table 4).
- The baseline comparisons are confounded:
- The task is identity reconstruction, where achievable fidelity is governed mainly by bottleneck capacity. Latent dimensions and compression ratios are not reported for any model.
- As described, the decoder upsamples only once, by a factor of two in-plane (Sec. 3.2, 3.2.3). This implies a latent with only four times fewer spatial positions than the input, which may explain the PSNR gap over the MONAI autoencoder and the ViT.
- All models are trained for only 20 epochs.
- The most relevant baselines are absent: The paper cites dynamic and conditional convolutions (CondConv, Dynamic Convolution, ODConv, KAConv) and spatial mixture-of-experts (Dryden & Hoefler, 2022), but compares against none of them.
Altogether, these make the significance of the contribution unclear.
2. The specialization claim is largely built in by design, and the manuscript contains internal inconsistencies.
- Differentiation is imposed architecturally: The experts differ by construction. In the decoder, E0 and E1 operate on low-pass and high-pass components respectively, and the others differ in kernel orientation (1×3×3 vs. 3×1×1). Distinct structural and frequency responses are therefore expected from the architecture and are not evidence of emergent functional specialization. Conversely, Table 1 lists encoder experts E1 and E4 both as 1×1×1 convolutions, so the bank is not fully heterogeneous.
- The routed-correlation analysis is mechanical: The routed contribution is defined as . Changes in its correlation with structural maps follow directly from multiplying by a spatially varying weight. They do not show that the router selects experts based on image structure.
- Expert roles are not reproducible: The paper reports that an expert's structural association in one run need not hold in another (Appendix C.7). Only aggregate redundancy is stable (0.58–0.63), so specialization at the level of individual experts is not established.
- The formulation is close to existing operators: As shown in Appendix A (Eq. 12), the operation reduces to a spatially varying kernel . This is a spatially varying form of dynamic convolution, closely related to the cited KAConv, pixel-adaptive convolution, and spatial mixture-of-experts.
- The manuscript contains internal inconsistencies:
- The same model is reported with MSE 0.00147 and SSIM 0.955 in Table 3, but MSE 0.001310 and SSIM 0.9596 in Table 5, at identical PSNR (29.04 dB).
- Table 1 lists decoder E3 as 3×1×1 and E4 as 1×1×1; Algorithm 2 assigns them the reverse.
- The encoder is described as having three blocks (Sec. 3) and as two parts of two blocks each (Sec. 3.1).
- Algorithm 1 subtracts the spatial mean of the routing logits, whereas Eq. 20 subtracts the mean across experts.
- Figure 13 is captioned as model reconstructions but shows routing maps, and complexity results are attributed to Table 2 rather than Table 4.
- Template placeholder text remains in the Author Contributions and Acknowledgments sections.
- Channel widths, latent size, window size, and the number of attention heads are not reported.
Questions for authors
- Bottleneck capacity: What are the latent dimensions and compression ratios of the proposed model, the MONAI autoencoder, and the ViT? How do the models compare at matched compression?
- Value of routing: Given that a single unrouted expert (E2) matches the full adaptive model, what does spatial routing contribute? Please report mean ± standard deviation over seeds for the adaptive, uniform, and single-expert configurations.
- Related operators: How does the method compare with CondConv, Dynamic Convolution, ODConv, and spatial mixture-of-experts at matched parameter count and compute?
- Inconsistencies: Please reconcile the reconstruction metrics in Tables 3 and 5, the decoder expert definitions in Table 1 and Algorithm 2, and the logit normalization in Algorithm 1 and Eq. 20.
- Downstream relevance: Does the learned expert composition provide any benefit on a task other than identity reconstruction, such as segmentation or super-resolution?
Clarification of Routing Contributions, Model Comparisons, and Scope
We thank the reviewer for the careful assessment and for recognising the uniform-aggregation control and the qualified treatment of our analyses. We agree that the results warrant a narrower interpretation and address the five questions below.
1. Bottleneck capacity and reconstruction comparisons
We agree that comparisons with MONAI and the 3D ViT are difficult to interpret without reporting latent dimensions and compression ratios. The decoder's 2× in-plane upsampling alone does not establish the overall compression ratio. Further, training all models for 20 epochs does not guarantee convergence. We therefore interpret these results only under the reported training setup, not as capacity-controlled evidence of architectural superiority.
Our task is volumetric MRI autoencoding: fully observed IXI T1-weighted volumes serve as both inputs and targets. It does not model undersampled acquisition or k-space data consistency, and we make no claims of clinical utility. We use it as a testbed for studying spatially routed heterogeneous 3D operators on structured volumetric data.
2. Value of spatial routing
We acknowledge that adaptive routing yields a modest gain over uniform aggregation: 0.20 dB PSNR and 0.0010 SSIM. E2-only is also competitive (29.14 dB, 0.9593 SSIM), while removing E0 or E3 improves reconstruction fidelity. Thus, the results do not establish that the full five-expert bank is optimal or that routing is necessary for this task. However, removing E2 substantially degrades the full model (24.87 dB), indicating its importance within the bank. A single expert can achieve competitive fidelity, but cannot exhibit spatial composition among multiple operators. Our focus is on characterising that composition, not claiming consistent reconstruction superiority.
We agree that mean ± standard deviation across seeds for adaptive, uniform, and E2-only configurations would strengthen the comparison. Appendix C.7 examines variability in expert-level routing patterns and structural associations; it does not establish the statistical significance of the reconstruction gain. We therefore make no such claim.
We also acknowledge the computational overhead relative to MONAI (347.79 vs. 26.35 ms inference; 2.276 vs. 0.313 GB peak memory). We describe the model as parameter-efficient, not computationally efficient. Although high normalized entropy (0.957) indicates diffuse expert weighting rather than selective routing, entropy alone does not establish spatial uniformity. The narrow routing ranges nevertheless indicate relatively modest, soft modulation, and we have qualified our description accordingly.
3. Relationship to related operators
We agree that our formulation is closely related to spatially varying dynamic convolution and mixture-of-experts approaches. Appendix A expresses the weighted operation as a spatially varying effective kernel; we do not claim this general formulation as fundamentally new. Our focus is on characterising the spatial contributions of heterogeneous 3D operators and their associations with structural and frequency properties. We have not performed matched-parameter and matched-compute comparisons against CondConv, Dynamic Convolution, ODConv, or spatial MoE, and do not claim superiority over them.
4. Architectural interpretation and inconsistencies
The experts' different operators impose distinct inductive biases by design. Our claim is not that routing creates these properties or discovers fixed semantic roles, but that it modulates their contributions spatially. Since Ck=wkAk, correlations involving routed contributions can also be affected mechanically by the weights. We therefore treat these as descriptive associations, not evidence that the router independently selects experts based on image structure. Appendix C.7 also shows that individual expert associations vary across seeds; we do not claim reproducible expert-specific roles.
We have corrected the inconsistencies identified by the reviewer: reconstruction metrics in Tables 3 and 5, decoder expert definitions in Table 1 and Algorithm 2, and logit normalization in Algorithm 1 and Eq. 20. We have also corrected the encoder-block description, Figure 13 caption, complexity-table reference, and template placeholders.
5. Downstream relevance
We agree that identity autoencoding cannot establish downstream utility. We have not evaluated segmentation or super-resolution and make no claims of improvement on these tasks. Evaluating transfer to downstream volumetric tasks remains future work.
Overall, we have narrowed the framing to distinguish the analysis of spatially modulated operator contributions from claims of substantial reconstruction gains, universally necessary experts, or downstream utility.