- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
TIGR-Seg combines Gaussian-Gated Feature Adaptation (GGFA) of a frozen Swin encoder with a decoder that refines text-initialized spatial priors using image features. It reports Dice scores of 92.78%, 84.76%, and 86.77% on BKAI, BTMRI, and BUSI, respectively, with 4.6M trainable parameters. The experiments are promising, but the methodological advance and reliability of the text-refinement mechanism remain insufficiently established.
Critical strengths
- Coherent architecture and motivation. Combining parameter-efficient visual adaptation with text-guided spatial refinement is sensible. The encoder and decoder components have clear, complementary purposes.
- Encouraging adaptation results. Supplementary Table 6 compares GGFA with LoRA, Houlsby Adapter, FiLM, and full fine-tuning while retaining the remaining architecture. GGFA achieves the highest reported Dice/mIoU on both evaluated datasets with 4.6M trainable parameters.
- Interesting decoder formulation and useful ablations. Treating text as an initial hypothesis is a reasonable design. The bounded covariance parameterization is valid, and the paper investigates Gaussian gating, correlation, decoder components, prompt sensitivity, and cross-dataset generalization.
Critical weaknesses
- Moderate novelty and incomplete attribution. GGFA combines context-conditioned affine modulation with Gaussian gating, while TIGR adds image-conditioned corrections to a text-derived Gaussian prior. The combination is reasonable, but the conceptual advance over existing adaptation and multimodal conditioning is limited. Although the supplementary ablations are useful, there is no targeted comparison with fixed or image-only Gaussian initialization to establish why text initialization itself is necessary, rather than Gaussian conditioning more generally.
- Prompt robustness is weaker than the framing suggests. In supplementary Table 5, removing text reduces Dice by 13.30 points on BKAI and 15.51 on BUSI; contradictory locations reduce it by 16.44 and 20.85 points. Even a longer prompt with the same correct information causes noticeable degradation. These results demonstrate strong text dependence, but do not establish that image-guided correction can reliably recover from inaccurate guidance. The paper should distinguish successful refinement of accurate prompts from robustness to unreliable prompts.
- The equations do not fully support the interpretation of presence and reliability. Substituting Eq. (26) into Eq. (27) cancels the scalar presence score from the normalized weighted average. Thus, presence cannot directly attenuate that aggregation, although it can still affect reconstruction and gating in Eqs. (29) and (31). Also, the sigmoid correction weight in Eq. (19) is called reliability without calibration or an ablation against a constant weight. The roles attributed to these terms need more precise explanation and targeted validation.
- Evaluation independence and text provenance need clarification. Section 3.1 specifies fixed 80/20 splits but does not establish patient-disjoint splitting, particularly important for the brain MRI evaluation. It cites sources for structured descriptions without explaining whether test-time location and shape annotations require ground-truth masks or how equivalent guidance would be obtained in practice. This is not proof of leakage, but it leaves the independence and deployment relevance of the evaluation unclear.
- Statistical and ablation evidence is incomplete. Only TIGR-Seg has standard deviations in the main comparison, while competing methods and most ablations are reported as single values. The uncertainty of the comparative gains is therefore unclear. Table 2 omits the configuration with neither GGFA nor TIGR. The PEFT comparison also lacks explicit adapter/rank settings and tuning details, making it difficult to assess whether alternative adaptation methods received equally competitive configurations.
- Training efficiency does not establish a compact inference model. Supplementary Table 11 reports 166.9M total parameters, a 667.9 MB model, and 245.4 ms/image latency, despite only 4.6M trainable parameters. Several listed methods are smaller or faster. The authors acknowledge this distinction, but the broader efficiency framing should consistently emphasize parameter-efficient training. Timing details such as batch size, precision, warm-up, and inclusion of text encoding are needed to interpret the latency comparison.
Questions for authors
The architecture is coherent, the reported segmentation performance is strong, and the supplementary experiments provide useful evidence. However, the novelty is moderate, the central correction mechanism remains highly dependent on prompt quality, and important aspects of evaluation independence, comparative uncertainty, and mathematical interpretation need clarification.