- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes AAMamba, a compact Mamba-based model for 3D medical image segmentation. It argues that how a volume is turned into a token sequence is a main bottleneck. A content-adaptive tokenizer warps the sampling grid toward regions with strong intensity change or edges or curvature. The adaptive tokens are ordered with a bidirectional Hilbert curve and a training-time ROI mask boosts foreground tokens. A decoder then fuses the adaptive features with a grid-aligned stream. With 5.7M parameters it reports 84.2 mean Dice on BraTS and 91.85 on ACDC and 84.26 on Synapse and beats SegFormer3D on all three.
Critical strengths
- I like the question the paper asks. How to turn a 3D volume into a 1D sequence for Mamba is a real design issue and most medical Mamba work just uses a fixed scan. The density map and the dual stream idea are simple and easy to understand.
- The model is small at 5.7M parameters and still improves over SegFormer3D on every dataset. The gains are largest on small structures like ET and gallbladder and pancreas which fits the motivation.
- The ablation in Table 4 covers CAT and the grid stream and Hilbert order on all three datasets. It also shows that CAT matters most and that Hilbert order adds only a little.
Critical weaknesses
- The central claim that sequence construction matters as much as the sequence model is not supported by the ablation. Removing Hilbert in favor of raster costs only 0.2 Dice on BraTS and 0.14 on Synapse and 0.13 on ACDC. With one run and no error bars this is within noise. Most of the gain comes from CAT which is a token allocation idea and not a scan order idea.
- The ROI guidance does not match what the paper says. The text and Figure 4 describe lesion-first scanning but Eq. 2 only multiplies tokens by 1 plus alpha times the ROI value. Nothing is reordered. The ROI is also taken from the ground truth mask during training and set to zero at test time so the model sees a different input distribution at inference. There is no ablation of alpha or of removing ROI guidance.
- The tokenizer is under-specified. The displacement field Delta and the gate m are never defined. It is unclear whether they are learned or hand made and how fold-overs are prevented. Table 5 talks about control points P which never appear in the method section.
- The experimental comparison is weak. Only one Mamba model appears and only on BraTS and the numbers look copied from other papers. The BraTS and Synapse baselines such as TransUNet at 64.4 on BraTS are far below what is usually reported. There is no nnU-Net or U-Mamba or VM-UNet and no matched retraining. SegFormer3D has 4.5M parameters and AAMamba has 5.7M so the gap is not purely free. Only Dice is reported with no HD95 and no FLOPs or latency and no significance test even though the paper claims efficiency and thin structure gains.
Questions for authors
- How exactly are Delta and m computed and are they learned? What are the control points P in Table 5
- Does ROI guidance reorder tokens or only scale them? Please add an ablation over alpha including alpha equal to zero
- Were all baselines retrained under the same protocol and splits? Please add nnU-Net and U-Mamba or SegMamba on all three datasets.
- Can you report mean and standard deviation over several seeds and HD95 and also FLOPs and memory and latency with and without CAT and Hilbert?