- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes HyperGen, a framework for controllable pathological anomaly synthesis based on hyperbolic semantic manifold learning. The method first extracts lesion-specific representations through mask-guided feature aggregation and maps them into a Poincaré ball, where the radial coordinate is explicitly aligned with lesion attenuation (HU values), while angular directions are intended to capture morphological and textural variations. The learned representations are then mapped back to the tangent space and injected into a 3D diffusion model through FiLM conditioning, combined with step-wise background stitching to preserve surrounding anatomy. By manipulating the radial and angular coordinates, the authors aim to achieve interpretable control over synthesized lesions.
The method is evaluated on LIDC-IDRI and EMIDEC. The authors report improvements in generation quality measured by FID, SSIM, and PSNR, as well as enhanced downstream nnU-Net segmentation performance using synthetic data augmentation. Qualitative visualizations are further provided to illustrate semantic controllability and pathological condition preservation.
Critical strengths
1. Clinically meaningful semantic conditioning.
The paper addresses an important problem in controllable medical image synthesis by incorporating clinically interpretable attributes into the generation process. In particular, aligning the latent radial coordinate with lesion attenuation (HU values) provides an intuitive way to control pathological appearance. This is a potentially useful direction for generating diverse lesions with meaningful radiological characteristics, although the necessity of hyperbolic geometry remains unclear.
2. Evaluation of downstream clinical utility.
Beyond conventional image quality metrics, the paper evaluates the usefulness of synthesized lesions for downstream segmentation. The reported improvements in nnU-Net performance (e.g., Dice from 78.26% to 83.02% in the normal-case augmentation setting) suggest that the generated data may provide practical benefits for medical image analysis. This is a positive aspect of the evaluation, although stronger baseline comparisons and controlled ablations are still needed to establish the source of these gains.
Critical weaknesses
1. The necessity and contribution of hyperbolic geometry are not convincingly established.
The central methodological claim is insufficiently justified. According to Eqs. (4) and (7), the hyperbolic embedding is mapped back to the tangent space through the inverse logarithmic map, yielding the original Euclidean feature vector in the absence of navigation. Moreover, the proposed radial navigation (Eqs. (10)–(11)) is mathematically equivalent to scaling the Euclidean feature vector by a scalar factor. The HU-based radial alignment can also be formulated as a constraint on the Euclidean feature norm. Although hyperbolic geometry may influence the representation learning stage, the paper does not demonstrate any advantage over an equivalent Euclidean formulation. Without a carefully matched Euclidean baseline, it is difficult to determine whether the reported improvements arise from hyperbolic geometry or simply from HU supervision and feature scaling. This substantially weakens the paper's central novelty claim.
2. Insufficient experimental evidence to support the claimed controllability and performance advantages.
The experimental evaluation lacks several critical controls. First, no systematic ablation isolates the contributions of hyperbolic geometry, HU-based supervision, and manifold navigation. Second, the claimed disentanglement between lesion density and morphology is primarily supported by qualitative visualizations, without quantitative measurements of control accuracy, monotonicity, or attribute independence. Third, the improvements over strong lesion-focused diffusion baselines are relatively modest in some settings (e.g., 83.02 vs. 82.66 Dice), yet no variance estimates or statistical significance tests are reported. The baseline comparisons and cross-modality evaluation protocols also require further clarification. Consequently, the current experiments do not provide sufficient evidence to support the claimed geometric advantages or the generality of the proposed approach.
Questions for authors
Q1. Necessity of Hyperbolic Geometry (Clarification and Essential Validation)
According to Eqs. (4) and (7), the composition of the exponential and logarithmic maps recovers the original Euclidean tangent vector. Furthermore, the proposed radial navigation appears mathematically equivalent to scaling the Euclidean feature vector by a factor of . Could the authors clarify which component specifically benefits from non-Euclidean geometry? More importantly, can the authors provide a controlled comparison against an equivalent Euclidean formulation with the same HU supervision, architecture, and radial scaling? This is essential for assessing the paper's central methodological contribution.
Q2. Baseline Comparisons and Experimental Provenance (Clarification)
Several baseline results in Tables 2 and 3 appear identical to those previously reported in LeFusion. Were these results independently reproduced or directly adopted from the original publication? If adopted, could the authors clarify the sources and explain the differences in baseline citation attribution? Additionally, why are stronger published LeFusion configurations, particularly LeFusion-J on EMIDEC, not included in the comparisons? A clarification of the experimental protocols and comparison fairness would materially affect my assessment of the claimed performance improvements.
Q3. Quantitative Validation of Semantic Controllability (Essential Validation)
The paper claims that radial navigation controls lesion attenuation while angular navigation controls morphology and texture. However, the evidence is primarily qualitative. Can the authors provide quantitative evidence showing (i) the relationship between radial scaling and the actual HU values of generated lesions, and (ii) whether angular manipulation changes morphological characteristics while preserving lesion attenuation? Without such evidence, it is difficult to assess whether the proposed representation genuinely supports disentangled semantic control.
Q4. Cross-Modality Evaluation on EMIDEC (Clarification)
The proposed radial alignment relies on CT Hounsfield Units, whereas EMIDEC contains cardiac MRI data. Could the authors clarify how the HU-based supervision is adapted to MRI, and whether the model is retrained on EMIDEC or directly transferred from LIDC-IDRI? This distinction is important for evaluating the claimed generalizability of the learned representation. Additionally, could the authors clarify how the average metrics in Table 3 were calculated, as some reported averages do not match the arithmetic means of the corresponding subsets?