- Rating
- 3(+1)
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates OCT-to-OCTA reconstruction and proposes a one-step generative framework that uses deterministic translation as a reference. The method first trains a reference network with paired supervision, then learns corrections using a pMF network initialized to produce zero output. Through a local theoretical analysis, the paper relates reference prediction errors to transport corrections and sampling sensitivity. The approach is evaluated on OCTA-3M, OCTA-6M, and OCTA2024 through reconstruction comparisons, ablation studies, and mechanistic analyses. The results demonstrate competitive reconstruction performance and improvements in the main MAE metrics over direct pMF.
Critical strengths
(1) Competitive reconstruction results provide empirical support for the method’s practical value. The method achieves competitive volumetric and projection reconstruction results across multiple datasets without requiring additional vessel or retinal-layer annotations. In Table 3, both the frozen-reference and fine-tuned-reference variants improve the main MAE metrics over direct pMF, providing initial evidence for the effectiveness of the reference strategy under different training settings. These results are the main empirical consideration supporting acceptance. (2) The method has a clear optimization motivation supported by theoretical and mechanistic analyses. The reference prediction, zero-output initialization, and transport correction form a coherent design. The supplementary material further examines local properties under both frozen-reference training and joint fine-tuning, and relates sampling sensitivity to output variance. These analyses offer a plausible explanation for why a reference prediction may reduce the correction burden and make the proposed mechanism amenable to empirical verification.
Critical weaknesses
(1) The distinct novelty and practical value of the generative correction are not sufficiently established. The first contribution characterizes the task as “strongly paired yet underdetermined.” This characterization helps motivate the study, but the new insight established by presenting it as an independent contribution remains unclear. Moreover, prior work such as ResDiff and PMRF has explored constructing a deterministic prediction followed by generative refinement or transport. The paper therefore needs to distinguish more clearly its specific technical advances in pMF parameterization and training. In addition, reductions in transport corrections and sampling variability primarily characterize internal model behavior; they do not by themselves establish better recovery of fine vessels or structural continuity. These gaps limit confidence in the novelty and practical significance of the contribution. (2) Missing controls and statistical validation make the source and stability of the improvements difficult to determine. Compared with direct pMF, the full method introduces a wider reference network and an additional pretraining stage, simultaneously changing model capacity, initialization, and training resources. The absence of reference-only predictions, deterministic refinement baselines, and comparisons with comparable capacity and training budgets makes it difficult to isolate the benefit of the generative correction. Furthermore, the experiments use a single training seed and do not report confidence intervals or significance analyses for the main performance differences. Repeated noise sampling from a fixed trained model cannot substitute for independent training runs. Consequently, the current evidence does not sufficiently establish that the observed gains are reproducible and primarily attributable to the proposed core design.
Questions for authors
(1) Relative to methods such as ResDiff and PMRF that perform prediction followed by refinement or transport, what are the specific new contributions in the proposed pMF parameterization and training mechanism? What verifiable insight does the “strongly paired yet underdetermined” characterization provide beyond motivating the approach? (2) Could the authors report the performance of the reference network alone and a deterministic refinement model, together with comparisons using model capacity and training budgets that are as comparable as practicable? Do these results continue to support an independent benefit from generative correction and the claimed improvements in fine-vessel recovery and structural continuity? (3) Could the authors provide subject-level paired confidence intervals or appropriate significance tests for the main performance differences in Table 3, along with results from multiple independent training seeds? This evidence would directly affect the assessment of the stability and practical significance of the reported improvements.