{"id":"ca17f2a8-e356-46e4-a0c8-a1dfc84a00fe","arxiv_id":"2508.09645","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PG-SAM pairs expert diagnostic text with a Segment Anything Model to segment parotid lesions in multi-sequence MRI, reporting the best DSC in most comparisons across three hospital datasets, but the text may leak ground-truth size and location.","lead":"This paper presents PG-SAM, a model that uses radiologists' written diagnosis reports to guide SAM in outlining parotid gland lesions on MRI scans. It reports improvements across three hospitals, but the reports contain each lesion's size and location, which may give the model an unfair advantage over comparison methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PG-SAM's expert text includes lesion size/location from the same target used to define GT; the SOTA comparison in Table 2 is therefore confounded by privileged input.","rationale":"The reader's weakest assumption is exactly the leakage issue, and my independent reading converges on the same point. Section 3.4 describes how the expert text is encoded and converted into a coarse mask, from which the prompt point and bounding box are derived. The text template in Table 4 explicitly includes [pos] and [xx] filled with the lesion's location and size. Since the ground-truth mask is the same lesion, the text carries target-derived information that the baselines never receive. This is a textbook privileged-information confound: the Table 2 comparison against text-free methods cannot support the stated SOTA claim, because the observed gains could come from the text rather than from the cross-sequence attention or decoder design. The paper even acknowledges the analogous leakage problem for GT-derived prompts in §3.4, which strengthens the concern. I considered the co-registration issue raised by the reader, but it is secondary: even if the sequences were perfectly aligned, the leakage confound would remain. The proposed concrete test would settle the matter by redacting the target-derived fields and by giving a baseline equivalent prompts. I do not recommend changing the reader's CONDITIONAL verdict: the concern is real, but it can be addressed empirically, and the method may still be valuable in settings where such diagnostic text is available at inference time.","tokens_in":15466,"tokens_out":5009,"duration_ms":52588,"concrete_test":"Re-run the Table 2 evaluation with two changes: (1) replace the expert text with a redacted version that removes all size and location information (e.g., 'There is a parotid gland lesion') for PG-SAM; (2) give Med-SA the same prompt points/bboxes that PG-SAM derives from the full expert text. If redacted-text PG-SAM falls to roughly the 'None' row of Table 4 or no longer beats Med-SA with equivalent prompts, then the Table 2 gains are due to target-derived text, not the proposed cross-sequence attention or decoder. A supporting check: compare PG-SAM using texts whose size/location fields are intentionally wrong; if DSC drops sharply, the model is relying on those fields.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim is confounded by target-derived privileged input. In §3.4 and Table 4, the expert diagnosis text uses the template 'This is an MRI image with a lesion located in the [pos] parotid gland, with size of [xx] mm.' The placeholders are filled with the actual location and dimensions of the lesion (Fig. 2 shows '29 × 25 × 35 mm'). That lesion is exactly what defines the ground-truth mask used for evaluation. The text is encoded with MedCLIP, fused with image features, and used by the prompt-generation module to produce the coarse mask and the prompt point/bbox that condition SAM's decoder (§3.4). The compared methods (UNet, UNet++, UNeXt, SwinUNet, TransUNet, EMCAD, Med-SA) receive no such target description, so their lower DSC in Table 2 may reflect missing privileged input rather than inferior architecture. The paper itself warns against this style of leakage in §3.4 ('Early approaches extracted prompts directly from ground truth, which led to label leakage'), but the diagnostic text's location/size fields are target-derived measurements; the coarse mask and bbox are literally computed from text-derived priors. This does not refute PG-SAM as a text-conditional segmenter, but it does invalidate 'state-of-the-art performance' against text-free baselines. A secondary issue is the unstated assumption in §3.3 that the four sequences are co-registered; no alignment step is described.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PG-SAM, a SAM-based framework for multi-sequence MRI segmentation of parotid gland lesions. A prompt-generation module uses expert diagnostic text (encoded with MedCLIP and adapted) together with multi-sequence image features to produce a coarse mask, from which point and bounding-box prompts are derived. A cross-sequence attention module fuses T1/T1C/T2/ADC embeddings, and a hierarchical mask decoder (with class-balanced mask-guided attention and learnable mask cross-attention) produces the final segmentation. The method is trained on Site1 and evaluated on internal Site1 and external Site2/Site3 datasets, reporting DSC, HD95, ACC, and REC against UNet-family, transformer, and SAM-based baselines. The authors claim state-of-the-art performance across three centers.","tokens_in":15824,"tokens_out":3612,"duration_ms":42919,"significance":"If the central claim were established, the paper would make a useful contribution to SAM-based medical image segmentation by showing how expert text can replace manual prompts. The proposed architecture is plausible and the reported internal results are competitive. However, the evaluation is confounded: the expert text explicitly contains the lesion's location and dimensions, which are measurements of the same target that defines the ground-truth mask. The baselines receive no such privileged input, so the Table 2 comparison does not demonstrate architectural superiority. External-site gains are small and sometimes negative, no error bars or significance tests are reported, and the ablation tables do not consistently support the claimed benefits of the two main modules. These issues are load-bearing for the paper's main claims; with a fair comparison and proper statistical analysis the method may still be valuable, but the current manuscript does not establish it.","major_comments":[{"comment":"The expert text template in Table 4 is 'This is an MRI image with a lesion located in the [pos] parotid gland, with size of [xx] mm.' The placeholders are filled with the actual location and size of the lesion (e.g., '29 × 25 × 35 mm' in Fig. 2). This information is derived from the same lesion that defines the ground-truth mask used for evaluation. Section 3.4 then uses the text embedding to generate a coarse mask and, from it, the prompt point and bounding box. Thus PG-SAM receives target-derived privileged input, whereas all baselines in Table 2 are text-free. The comparison therefore confounds method quality with input information and does not support 'state-of-the-art performance' against those baselines. Please add controlled comparisons: (i) baselines given equivalent ground-truth-derived bounding-box/point prompts, (ii) PG-SAM with text stripped of location/size fields, and (iii)","section":"§3.4, Table 4, Fig. 2"},{"comment":"The external-site results do not consistently support the SOTA claim. On Site2 T1, PG-SAM scores 0.676 DSC versus 0.697 for Med-SA; on Site2 T2, 0.678 versus 0.637; on Site3 T1, 0.717 versus 0.704. The differences are small, and no standard deviations, confidence intervals, or significance tests are reported. With single-seed results, the claim of 'state-of-the-art performance across three independent clinical centers' is not statistically supported. Please report results over multiple runs and perform paired significance tests (e.g., paired bootstrap or Wilcoxon signed-rank over patients/slices).","section":"Table 2"},{"comment":"The ablation results do not consistently show that the proposed modules help. In Table 3, on Site2 T1C, adding TPM to CAM decreases DSC from 0.690 to 0.635; on Site1 T1, adding CAM increases from 0.760 to 0.765 only. In Table 4, expert text versus 'None' on Site2 T2 gives 0.650 vs 0.660, and on Site3 T1 0.666 vs 0.661. These differences are within the noise of a single run. The conclusion that expert text and cross-sequence attention 'significantly improve' segmentation needs statistical support and, if appropriate, a more careful per-modality analysis.","section":"Tables 3 and 4"},{"comment":"The cross-sequence attention module operates pixel-wise across the four MRI sequences without any description of registration or alignment. Since T1, T1C, T2, and ADC are acquired separately, they may not be voxel-aligned; Eq. (2) computes attention between spatial feature locations, which assumes a one-to-one spatial correspondence. Please state the registration/alignment procedure used in preprocessing, or validate that the sequences are co-registered in the datasets. Without this, the cross-sequence fusion results are difficult to interpret.","section":"§3.3, Eq. (2)"}],"minor_comments":[{"comment":"The Accuracy formula appears to be (TP + FN) / (TP + TN + FP + FN), which is not accuracy. It should be (TP + TN) / total. The text around the equation also mislabels TN and FN definitions.","section":"Eq. (9)"},{"comment":"There are numerous typos: 'Mathmatics', 'rencent', 'breifly', 'VMLs', 'donotes', 'adpater', 'H95' vs 'HD95' in Table 2, and inconsistent use of 'Inner Site'/'Inter Site'. Please proofread carefully.","section":"Throughout"},{"comment":"The caption says 'The first row to the second-last row' but should likely be 'the first row to the last row' for the methods; the ground truth is listed separately. Please clarify.","section":"Fig. 6 caption"},{"comment":"No code or data availability statement is included. Given the clinical dataset, at least a code release would aid reproducibility.","section":"§4.1.2"},{"comment":"The notation x_f is defined as both concatenated embeddings and fused embeddings with the same symbol. Please use distinct symbols (e.g., v_i and v_f) to avoid confusion.","section":"Fig. 3 and §3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope, and the idea of using expert diagnostic text to guide SAM is timely. However, the central SOTA claim is currently confounded by the privileged spatial information in the text prompt. I would not reject outright because the issue is addressable with additional controlled experiments and reframing, but the authors must either remove the target-derived fields from the text or include baselines with equivalent input information. Without that, the comparison is not informative. The weak external-site performance and lack of statistical testing further support major revision rather than acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real engineering effort with a valuable three-center dataset, but the headline claim doesn't survive contact with the protocol. The expert text in Table 4 includes the lesion's location and dimensions — measurements of the same lesion that defines the ground-truth mask — and the text-free baselines get none of that. So Table 2 is comparing a model that receives a strong spatial hint against models that don't.\n\nWhat's genuinely new: this is the first application of expert-report-guided prompt generation to parotid gland MRI segmentation, and the cross-sequence attention fusion is simple but sensible. The dataset itself (507/60/149 patients, four sequences, three centers) is a solid contribution. The internal ablation shows both modules help, which is credible.\n\nThe soft spots are serious. Section 3.4 explicitly warns that extracting prompts from ground truth leads to label leakage, then the method does the equivalent via text: the size and location fields are filled from the diagnostic report that describes the same lesion defining the GT. The coarse mask, bbox, and points are all computed from that text. On top of that, external-site gains are small and sometimes negative (Site2 T1: 0.676 vs Med-SA's 0.697), and there are no error bars or significance tests, so \"state-of-the-art across three independent centers\" is overstated. Less critical: the four MRI sequences are assumed co-registered without an alignment step, the slice-inclusion threshold (>25 lesion pixels) looks post hoc, and \"eliminating the need for manual annotations\" is wrong — they still train and evaluate on pixel-level labels.\n\nOne thing in the paper's favor: the ablation in Table 4 shows expert text with size/location barely beats generic prompts on several rows (Site1 T1: 0.765 vs 0.762; Site2 T1: 0.705 vs 0.700). That suggests the leaked measurements are not the main driver of the gains, which is worth acknowledging. But it doesn't fix the unfair comparison against text-free baselines.\n\nWho this is for: people working on SAM prompting for medical images or head-and-neck MRI segmentation. The dataset and the general idea are worth engaging with, and the paper deserves a real referee — but the referee should insist on matched text conditions (either give baselines the same text or evaluate with reports stripped of target measurements) and variance reporting. As it stands, the method is plausible as a text-conditioned segmenter, but the claimed superiority over prior art is not established.","headline":"A reasonable engineering paper with a useful dataset, but the central SOTA claim is confounded because the expert text carries target-derived location and size information that the baselines never see.","tokens_in":16321,"tokens_out":1724,"would_cite":false,"duration_ms":19260,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Structured expert diagnostic text, turned into point-and-box prompts, can guide a fine-tuned Segment Anything Model to segment parotid gland lesions across multiple MRI sequences and three clinical centers.","keywords":["parotid gland lesion segmentation","Segment Anything Model","expert diagnostic text","cross-sequence attention","multi-sequence MRI","prompt generation","medical image segmentation","LoRA fine-tuning"],"falsifier":"On the Site1 internal test set, rerun the trained PG-SAM with expert text edited so that the size field [xx] is replaced by a random value from the dataset's size distribution and the location field [pos] is moved to the contralateral side, keeping images, masks, and all other inputs fixed. If DSC and HD95 remain at the reported levels, target-derived text is not the source of the gains; if they collapse, the reported advantage over text-free baselines is partly an artifact of privileged ground-truth information in the prompt.","tokens_in":15360,"feed_emoji":"🎯","tokens_out":7181,"duration_ms":70617,"temperature":0.7,"pith_summary":"The paper proposes PG-SAM, a parotid-gland lesion segmentation model that uses structured expert diagnostic text—location and size of the lesion—to automatically generate the point and box prompts that SAM needs, instead of relying on a user to click on the lesion. It couples this prompt generator with a cross-sequence attention module that fuses four MRI sequences (T1, T1C, T2, ADC) into a shared embedding. On three clinical datasets, PG-SAM reports the best or second-best Dice similarity coefficient on most modality–center combinations, including internal and external test sets, with the largest gains on T1 and T1C sequences. The authors interpret this as evidence that radiological domain knowledge, encoded as text, improves SAM-based medical image segmentation without manual annotations.","feed_headline":"Expert text prompts improve parotid-lesion MRI segmentation","feed_subtitle":"A model turns location and size from radiology reports into SAM prompts and fuses four MRI sequences.","key_machinery":"The load-bearing machinery is the expert-text-to-prompt pathway: MedCLIP encodes the report, an adapter transfers the text embedding to the medical domain, a text guide decoder fuses it with the fused multi-sequence image embedding, and the resulting coarse mask is masked by a pre-computed spatial prior $M$ before a prompt point (centroid) and prompt box (maximal border) are extracted. These prompts are fed through SAM's prompt encoder into a hierarchical mask decoder with class-balanced mask-guided self-attention and learnable mask cross-attention. The cross-sequence attention formula $A_i = \\mathrm{Softmax}(x_f W_q (x_i W_k)^\\top / \\sqrt{d_k}) W_v$ is the mechanism that lets each sequence","core_discovery":"PG-SAM's central claim is that expert knowledge from diagnostic reports can replace manual prompts for SAM in parotid lesion segmentation. The expert text is embedded with MedCLIP, adapted to the medical domain, and fused with multi-sequence image features in a text guide decoder. A coarse mask is generated from that fusion, constrained by a spatial prior mask of where parotid lesions typically appear; the centroid and bounding box of the coarse mask become the prompt point and prompt box. Meanwhile, a cross-sequence attention module lets each sequence's embedding be corrected by information from the other sequences. The paper reports that this combined design outperforms the compared UNet-b","pith_inferences":["Because the template text includes the lesion's measured size (e.g., 29×25×35 mm) and its location, and because those numbers come from the same lesion that defines the ground-truth mask, part of the reported gain may be attributable to privileged target information rather than to the architecture; a blinded text condition would settle this.","A clean extension would randomize or swap only the size field while keeping location fixed, or vice versa, to identify which textual variable actually drives the prompt's value.","The cross-sequence attention fuses features pixel-wise without an explicit registration step; if clinical MRI sequences are misaligned, the benefit could flip to harm, so a test that shifts one sequence by a few pixels would quantify this sensitivity.","The same expert-text-to-prompt mechanism could transfer to other anatomically localized lesions (e.g., thyroid, lymph nodes) where structured radiology reports exist, but the spatial prior mask would need to be rebuilt for each anatomy."],"forward_implications":["If PG-SAM is correct, structured radiology reports can serve as cheap, automatically derived prompts for SAM, removing the need for per-slice manual point/box annotations.","Fusing T1, T1C, T2 and ADC through cross-sequence attention should make parotid segmentation more consistent across modalities than single-sequence models.","The paper's few-sample experiments imply that with only 10–30% of labeled slices, PG-SAM stays above 0.7 DSC on internal T1 and above 0.6 on external sets, suggesting the method is usable in low-annotation clinical settings.","The spatial prior mask, built from where parotid lesions occur, makes prompt generation more stable and is a component that could be reused for other anatomically localized lesions.","Combining text-guided prompts with parameter-efficient LoRA fine-tuning keeps the method practical: the paper reports training time cut to one-third and roughly 40% deployment-efficiency improvement relative to the traditional comparison methods."],"supporting_citations":[{"why":"Supplies the Segment Anything Model backbone: image encoder, prompt encoder, and mask decoder that PG-SAM adapts and fine-tunes.","marker":"[6]"},{"why":"MedCLIP is the pretrained vision-language encoder used to embed the expert diagnostic text into text embeddings.","marker":"[27]"},{"why":"SAMed contributes the LoRA-based fine-tuning strategy PG-SAM applies to the frozen SAM image encoder.","marker":"[14]"},{"why":"LoRA provides the low-rank update mechanism for parameter-efficient adaptation of the image encoder.","marker":"[30]"},{"why":"Medical SAM Adapter (Med-SA) is the strongest SAM-based baseline PG-SAM compares against and must beat.","marker":"[29]"},{"why":"SAM-SP is the self-prompting alternative that motivates PG-SAM's automatic prompt generation module.","marker":"[16]"},{"why":"H-SAM supplies the hybrid hierarchical mask decoder design (CMAttn and LMCA) used in the segmentation network.","marker":"[32]"},{"why":"Supplies the long-tail logit adjustment idea behind the class-balanced Gaussian noise in CMAttn.","marker":"[43]"},{"why":"Provides the adapter layer mechanism used to adapt the text embeddings to the medical domain.","marker":"[31]"}],"fun_headline_variants":["Expert text prompts improve parotid lesion MRI segmentation","Radiology report text guides SAM for parotid lesion segmentation","Text-guided prompts replace manual clicks in parotid MRI segmentation","Multi-sequence SAM with expert text improves parotid lesion accuracy","Diagnostic reports prompt SAM to better segment parotid lesions"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the expert diagnostic text used to build the prompts does not carry measurements (lesion size and location) taken from the same ground-truth masks the model is trained and scored against, and that the four MRI sequences are aligned well enough for pixel-wise cross-sequence fusion.","fun_headline_variants_meta":{"raw":{"variants":["Expert text prompts improve parotid lesion MRI segmentation","Radiology report text guides SAM for parotid lesion segmentation","Text-guided prompts replace manual clicks in parotid MRI segmentation","Multi-sequence SAM with expert text improves parotid lesion accuracy","Diagnostic reports prompt SAM to better segment parotid lesions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000914,"raw_usage":{"total_tokens":3771,"prompt_tokens":762,"completion_tokens":3009,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":2929}},"tokens_in":506,"tokens_out":3009,"duration_ms":23549,"temperature":1.0,"reasoning_tokens":2929,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:55:02.287152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the Site1 internal test set, rerun the trained PG-SAM with expert text edited so that the size field [xx] is replaced by a random value from the dataset's size distribution and the location field [pos] is moved to the contralateral side, keeping images, masks, and all other inputs fixed. If DSC and HD95 remain at the reported levels, target-derived text is not the source of the gains; if they collapse, the reported advantage over text-free baselines is partly an artifact of privileged ground-truth information in the prompt.","supporting_citations":[{"cited_title":"Kirillov, E","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model backbone: image encoder, prompt encoder, and mask decoder that PG-SAM adapts and fine-tunes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MedCLIP is the pretrained vision-language encoder used to embed the expert diagnostic text into text embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LoRA provides the low-rank update mechanism for parameter-efficient adaptation of the image encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Medical SAM Adapter (Med-SA) is the strongest SAM-based baseline PG-SAM compares against and must beat."},{"cited_title":"SAM-SP: Self-Prompting Makes SAM Great Again","cited_arxiv_id":"2408.12364","evidence_quote":"SAM-SP is the self-prompting alternative that motivates PG-SAM's automatic prompt generation module."},{"cited_title":"Cheng, Q","cited_arxiv_id":null,"evidence_quote":"H-SAM supplies the hybrid hierarchical mask decoder design (CMAttn and LMCA) used in the segmentation network."},{"cited_title":"Li, Y.-m","cited_arxiv_id":null,"evidence_quote":"Supplies the long-tail logit adjustment idea behind the class-balanced Gaussian noise in CMAttn."},{"cited_title":"Houlsby, A","cited_arxiv_id":null,"evidence_quote":"Provides the adapter layer mechanism used to adapt the text embeddings to the medical domain."}],"review_version":1}