REVIEW 4 major objections 6 minor
PromptForSegCXR: Prompt-Driven Multi-Organ and Multi-Disease Segmentation in Chest X-rays using a Multi-stage Fusion Mechanism
T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A doodle prompt can segment 6 organs and 17 diseases in chest X-rays with one light model.
desk verdict The first doodle-prompt multi-organ/multi-disease CXR dataset is a real resource, but the paper's headline outperformance claim is undermined by an apples-to-oranges comparison and a likely circular disease-mask provenance. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is a dual-branch encoder-decoder whose two branches process the chest X-ray and the doodle through DPRconvSE blocks, fusing their feature maps at five scales via concatenation. Each DPRconvSE block is a depthwise 3x3 convolution, pointwise 1x1 convolution, squeeze-and-excitation attention, and a residual connection. The decoder upsamples and concatenates with the fused features to produce a 256x256 mask, and the paper argues that this multi-stage fusion is what lets a lightweight model combine spatial detail with semantic context.
What would settle it
Run the same 23-class test split through Prompt2SegCXR with bounding-box prompts and through SAM, MedSAM, and SAM-Med2D fine-tuned on the same training split with doodle prompts; if any baseline reaches or exceeds the reported Dice of 81.62, the claim that the model outperforms state-of-the-art prompt-based segmentation is refuted.
Extended reading notes
Core claim
The paper claims that a model with only 8 million parameters can segment all 23 target classes, and that architecture choices, including depthwise separable convolutions, squeeze-and-excitation attention, residual connections, multi-stage feature fusion, and random oversampling, each contribute to the final score. Trained with 5-fold cross-validation on the curated dataset, the full model reports mean Dice 81.62, Jaccard 70.38, AUC 88.92, and accuracy 97.03; restricted to organs only it reports Dice 92.47, and restricted to diseases only it reports Dice 73.08. The comparison against existing prompt-based models reports higher Dice, Jaccard, AUC, and accuracy for Prompt2SegCXR. The paper also reports that the hardest classes are small or line-shaped disease regions such as Linear Density, Calcified Lymph Node, and Single Nodule.
Load-bearing premise
The central claim rests on comparing Prompt2SegCXR, which was trained on the target 23-class dataset and prompted with expert doodles, against SAM-family models used zero-shot with bounding-box prompts, and if that asymmetry is not a fair baseline, the outperformance claim does not follow.
Editorial extensions
If this is right
- A doodle prompt can specify any of 23 targets, organs and diseases, in a single model, so a user does not need a separate segmentation model per structure.
- With 8 million parameters, the architecture is light enough for resource-limited clinical settings while matching or beating much larger prompt-segmentation models on the reported metrics.
- The ablations imply that each component, depthwise separable convolution, SE attention, residual connections, and oversampling, contributes to the final Dice, so the recipe transfers to similar dual-input segmentation tasks.
- The released 23-class prompt dataset gives other researchers a benchmark for interactive chest X-ray segmentation.
Reading between the lines
- If the comparison were controlled so that SAM-family models were fine-tuned on the same training split, or Prompt2SegCXR were tested zero-shot, the reported margin might shrink or reverse; the paper's design does not isolate prompt modality from training status.
- The labor-intensive expert doodle generation suggests a practical extension: a model that synthesizes candidate doodles from clicks or text could make the approach easier to deploy.
- The low Dice scores for Linear Density, Calcified Lymph Node, and Single Nodule point to small, line-shaped targets as the bottleneck, and a higher-resolution branch or class-weighted loss is a direct testable improvement.
- The released prompt dataset, by pairing each chest X-ray with a prompt and mask, could be reused to train or evaluate other interactive segmentation models, including text-prompted ones, beyond this architecture.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents Prompt2SegCXR, a dual-input chest X-ray segmentation model that takes an image and a user-drawn doodle prompt as input and outputs a segmentation mask for one of 23 classes (6 organs, 17 diseases). The authors also introduce a doodle-prompt dataset assembled from JSRT, VinDr-RibCXR, and Shenzhen Hospital sources, with prompts created by medical experts. The model, built from depthwise/pointwise residual convolution blocks with Squeeze-and-Excitation attention and multi-stage feature fusion, reports an average Dice of 81.62, Jaccard of 70.38, AUC of 88.92, and accuracy of 97.03 on the full dataset, and it is compared against zero-shot SAM-family baselines that use bounding-box prompts.
Significance. If the dataset and results are valid, this would be a useful contribution: an interactive, lightweight prompt-based segmentation model for chest X-rays with a publicly released prompt dataset is a practical resource for clinical workflows and for benchmarking future methods. The paper also provides a detailed architecture and ablation study. However, the scientific claims rest on two load-bearing issues: the provenance of the disease ground-truth masks is not established, and the headline comparison against state-of-the-art methods is confounded by training status and prompt type. These issues currently prevent the results from being interpreted as a reliable demonstration of the claimed capabilities.
major comments (4)
- [§3.1–3.2] The disease ground-truth mask provenance is not established. Section 3.1 states that 'Disease-wise segmentation masks for the remaining classes were provided by this dataset' and cites the Shenzhen Hospital dataset [50], but this source is a tuberculosis classification collection without pixel-level segmentation annotations. The only annotation procedure described in Section 3.2 is that medical experts annotated CXRs to create doodle prompts and then 'refined' them by keeping the annotated pixel range. If the same expert annotations were used both as the doodle prompt and as the ground-truth mask, then the reported disease Dice scores (e.g., 38.69 for Calcified Lymph Node, 19.83 for Linear Density in Table 1) measure how well the model reproduces its input rather than how well it independently segments pathology. The authors must state explicitly how the disease masks were generated and whether they are independent of the prompt annotations; without this, the disease segmentation evaluation is potentially circular.
- [Table 5] The comparison in Table 5 is confounded. Prompt2SegCXR is trained on the target dataset with doodle prompts, while SAM, MedSAM, and SAM-Med2D are evaluated zero-shot with bounding-box prompts. Thus both training status and prompt type differ between the proposed model and the baselines, so the headline claim that the model 'outperformed state-of-the-art models' is not supported. A valid comparison would require the baselines to be fine-tuned on the same doodle-prompt dataset, or Prompt2SegCXR to be evaluated zero-shot on external data, or at minimum the same prompt modality for all models.
- [§5.1, Table 3] The disease-only results are reported selectively. Table 3 is titled 'excluding few' but does not specify which classes are excluded; the text says it excludes classes on which the model 'did not achieve attractive results.' From Tables 1 and 3, the excluded classes appear to include Linear Density, Thickening of Interlobar Fissure, and Single Nodule (non-calcified), which have some of the lowest Dice scores in Table 1 (19.83, 34.75, and 39.90, respectively). Excluding these low-performing classes inflates the reported disease-average Dice from 81.62 to 73.08? Actually the full model reports 81.62 over all classes; the disease-only model in Table 3 reports 73.08 over the selected subset. The comparison is not apples-to-apples, and the selective exclusion obscures the model's actual limitations on small or subtle disease regions. The authors should report all 17 disease classes, or clearly pre-specify inclusion criteria.
- [§4, §5] The results lack error bars and statistical significance testing. The text says all scores were averaged over three independent runs and that a 5-fold cross-validation was used, but no variance, confidence intervals, or significance tests are reported. Given the small per-class sample sizes (at most 50 per disease class) and the modest gaps in Table 5 (e.g., 81.62 vs. 74.82 Dice), the claim of superiority over the baselines is not statistically substantiated.
minor comments (6)
- [§3.4] The text refers to '64x64 filters' and '1024x1024' filter sizes in the encoder and decoder, which should be the number of feature maps (64, 128, 256, 512, 1024), not spatial dimensions. Please correct the wording.
- [§4.2] The relationship between the 5-fold cross-validation and the 'three independent runs' is unclear: please specify which numbers are averaged and whether the reported metrics are on the held-out test split.
- [Figure 1] Figure 1 is referenced but not visible in this version; the per-class counts of CXRs, masks, and prompts should be clearly displayed and also reported in a table.
- [Section 1] The claim of being the 'first doodle prompt-based multi-organ and multi-disease segmentation work' should be supported by a comparison with related interactive segmentation methods beyond SAM variants (e.g., click-based or scribble-based interactive segmentation).
- [Table 4] In Table 4, the entries '68.4592.4296.33' appear to have missing separators; please format the Jaccard, AUC, and Accuracy values clearly.
- [Conclusion] The conclusion repeats the claim that the model 'outperforms in both accuracy and precision' compared to state-of-the-art models, but this should be tempered given the confounded comparison in Table 5.
Circularity Check
The 17 disease-class 'predictions' may reduce to copying the doodle: the paper's only described annotation procedure generates a binary image from the expert-annotated pixel range, and no independent disease ground-truth mask source is actually described.
-
self definitional
[Sections 3.1-3.2 (Dataset Description and Prompt Preparation)]
"Disease-wise segmentation masks for the remaining classes were provided by this dataset. ... The doodle prompts were carefully crafted by the medical experts on our team. Using Microsoft Image Viewer, they annotated chest X-rays (CXRs) to highlight specific organs or disease-affected regions. The prompts were then refined by isolating the annotated pixel range, assigning them their intended color, and setting all other pixels to zero for clarity and precision."
For the 17 disease classes, the paper's only described mask-generation step is the Sec 3.2 prompt pipeline, which converts the expert annotation into a binary image by keeping exactly the annotated pixel range. No independent protocol for creating disease ground-truth masks is described, and no external disease-mask source with pixel annotations is actually identified. The same expert binary annotation therefore necessarily serves as both the doodle prompt (input) and the ground-truth mask (target). Under that setup, the reported disease Dice/Jaccard/AUC numbers, including the weak classes (e.g., Linear Density 19.83 Dice, Calcified Lymph Node 38.69 Dice), measure how well the decoder reconstructs its own input rather than how well it segments an independent ground truth.
full rationale
The organ-class part of the paper is not circular: lung, heart, clavicle, and rib masks come from independent external sources (Ginneken's SCR masks and VinDr-RibCXR), so those Dice scores are genuine segmentation results. The circularity is confined to the 17 disease classes, where the only described annotation step is the same expert annotation that produces the doodle prompt, and no independent disease-mask source is established. The model's disease-segmentation 'prediction' can therefore be satisfied by copying the input prompt. The Table 5 comparison with zero-shot SAM-family models is also confounded by training status and prompt modality (Ours is trained on the target dataset; SAM variants are used zero-shot with bounding boxes), but that is a benchmarking-validity concern rather than an equation-level circularity. Overall score 6: part of the central claim—multi-disease segmentation and the disease-specific metrics—reduces by construction, while the organ segmentation and architecture claims retain independent content.
Assumptions & free parameters
free parameters (3)
- Per-class training sample counts =
150 per JSRT organ, 100 VinDr ribs, up to 50 per Shenzhen disease
- Cropping padding =
100 pixels
- SE reduction ratio =
16
assumptions (4)
- domain assumption The Shenzhen Hospital dataset provides disease-wise segmentation masks.
- domain assumption Doodle prompts drawn by the author medical experts are a valid and sufficient prompt modality for interactive segmentation.
- domain assumption The radiograph and prompt can be fused by concatenation at five scales with no need for a learned prompt encoder.
- domain assumption The three source datasets can be pooled into one label space with 23 classes despite differing resolutions, scanners, and annotation conventions.
Cite this review
Pith. "Pith review of PromptForSegCXR: Prompt-Driven Multi-Organ and Multi-Disease Segmentation in Chest X-rays using a Multi-stage Fusion Mechanism." pith.science (2026). https://pith.science/paper/BN47ZBS2
@misc{pith2026250700673,
author = {Pith},
title = {Pith review of: PromptForSegCXR: Prompt-Driven Multi-Organ and Multi-Disease Segmentation in Chest X-rays using a Multi-stage Fusion Mechanism},
year = {2026},
howpublished = {\url{https://pith.science/paper/BN47ZBS2}},
note = {Machine review of arXiv:2507.00673}
}
read the original abstract
Image segmentation is central to automated medical image analysis, enabling precise identification of anatomical structures and pathological regions. Conventional segmentation models typically target a single organ or disease, limiting their adaptability across clinical scenarios. While multi-organ and multi-disease segmentation has been explored, building such datasets requires extensive manual annotation by medical experts. Prompt-driven segmentation offers a flexible, user-guided alternative that speeds up annotation, yet no prior work has addressed prompt-based interactive segmentation across multiple organs and diseases in chest X-rays. This study makes two main contributions. First, we introduce a novel dataset of expert-designed doodle prompts spanning 23 classes (six organs and seventeen diseases), curated from multiple public chest X-ray datasets for prompt-driven segmentation. Second, we propose PromptForSegCXR, a lightweight dual-input segmentation framework that combines the chest X-ray with user-provided doodle prompts to accurately segment diverse anatomical and pathological regions. The model uses a multi-stage feature fusion strategy to integrate spatial and semantic representations, along with a depthwise-pointwise-residual convolution block with squeeze-and-excitation attention for efficient hierarchical feature extraction and adaptive recalibration. Experimental results show the model achieves a Dice score of 81.62 percent on the full dataset, outperforming SAM-based prompt segmentation models by up to 10 percent and conventional segmentation architectures by up to 23 percent, while remaining lightweight. These results demonstrate the effectiveness of the proposed approach for accurate, flexible, prompt-driven chest X-ray segmentation.
Figures
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.