Eshragh, Mohammadmahdi, Mohammed, Emad A., Far, Behrouz, Weis, Ezekiel, Shields, Carol L, Ferenczy, Sandor R, Crump, Trafford · Sep 29, 2025
Paper authors propose hybrid method for fundus lesion segmentation
Markdown · JSON · Original source
The authors propose a hybrid pipeline for delineating choroidal nevi (CN), common benign fundus lesions that need longitudinal monitoring because a small fraction transform into melanoma. Their approach uses an unmodified UNet trained on 128 * 128 images to estimate a coarse location, then applies SLIC, a training-free superpixel method, to recover lesion boundaries at full input resolution. The location prior also selects the superpixel count for each image automatically, avoiding manual tuning. On 25 held-out 1024 * 1024 images, the authors report a mean Dice of 90.9%; on 90 external images from a second centre, they report 76.8%, compared with 69.3% for nnU-Net. They say ablation studies identify automated superpixel count as the key component. The proposed method requires no and takes 1.6 seconds per 1024 * 1024 image on a laptop, according to the abstract. Escape Meridian’s interpretation is that the results suggest a way to combine a coarse learned location with boundary recovery at full resolution, rather than training a segmentation model directly at that resolution. The evidence described here is image-set performance; it does not establish effects on diagnosis, monitoring decisions or patient outcomes.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Sep 29, 2025 (2025-09-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T21:03:39.412Z. First observed: 2026-11-06T05:12:38.116Z. Retrieved: 2026-11-08T21:02:36.298Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported performance is based on 25 held-out images and 90 external images from a second centre; the supplied abstract does not establish how the results generalize beyond those image sets. The supplied abstract reports segmentation metrics and inference , not effects on diagnosis, monitoring decisions or patient outcomes.
Open question
Does the reported advantage persist beyond the 25 held-out images and 90 external images described in the abstract?
Source-stated
The authors propose a hybrid pipeline that uses an unmodified UNet for a coarse location prior and SLIC to recover lesion boundaries at full input resolution.
an unmodified UNet trained on 128 * 128 images provides a coarse location prior, while a training-free superpixel method (SLIC) recovers the boundary at the full input resolution.
Source-stated
The location prior automatically selects the superpixel count for each image, avoiding manual tuning.
The prior further drives automatic per-image selection of the superpixel count via an image-to-lesion ratio, eliminating manual tuning.
Source-stated
On 25 held-out 1024 * 1024 images, the proposed method achieved a mean Dice of 90.9%.
On 25 held-out 1024 * 1024 images, the proposed method achieved a mean Dice of 90.9% (IoU 87.9%)
Source-stated
On 90 external images from a second centre, the paper reports 76.8% for the proposed method and 69.3% for nnU-Net.
On 90 external images from a second centre, the advantage persisted (76.8% vs. 69.3% for nnU-Net).
Source-stated
The abstract reports that inference requires no GPU and takes 1.6 seconds per 1024 * 1024 image on a laptop.
Inference needs no GPU, requiring 1.6 s per 1024 * 1024 image
Escape Meridian interpretation
The supplied abstract reports image-set segmentation results but does not report effects on diagnosis, monitoring decisions or patient outcomes.
Accurate delineation at native resolution is challenging: CN typically occupy
Source and contributor context
Eshragh, Mohammadmahdi, Mohammed, Emad A., Far, Behrouz, Weis, Ezekiel, Shields, Carol L, Ferenczy, Sandor R, Crump, Trafford · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: cbc5b8b9-0acf-417b-876f-8592da2e5773.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T21:03:39.412Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: eb4825712b5b35ac3e37414ba3419f7c23fd6cc1d49c2fa8c42451af83865304. Source text hash: 6241951b9451f1a392607e7a574046dd001a32a087a8364c29a361b2ebea432f.
Revision and source-check history
Version 1 · 2026-11-08T21:03:39.412Z
Paper authors propose hybrid method for fundus lesion segmentation
The authors propose a hybrid pipeline for delineating choroidal nevi (CN), common benign fundus lesions that need longitudinal monitoring because a small fraction transform into melanoma. Their approach uses an unmodified UNet trained on 128 * 128 images to estimate a coarse location, then applies SLIC, a training-free superpixel method, to recover lesion boundaries at full input resolution. The location prior also selects the superpixel count for each image automatically, avoiding manual tuning. On 25 held-out 1024 * 1024 images, the authors report a mean Dice of 90.9%; on 90 external images from a second centre, they report 76.8%, compared with 69.3% for nnU-Net. They say ablation studies identify automated superpixel count as the key component. The proposed method requires no and takes 1.6 seconds per 1024 * 1024 image on a laptop, according to the abstract. Escape Meridian’s interpretation is that the results suggest a way to combine a coarse learned location with boundary recovery at full resolution, rather than training a segmentation model directly at that resolution. The evidence described here is image-set performance; it does not establish effects on diagnosis, monitoring decisions or patient outcomes.
Original source
cs.AI · Article
Source details
Original source date: Sep 29, 2025 (2025-09-29).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T21:03:39.412Z. First observed: 2026-11-06T05:12:38.116Z. Retrieved: 2026-11-08T21:02:36.298Z.
Limits
The reported performance is based on 25 held-out images and 90 external images from a second centre; the supplied abstract does not establish how the results generalize beyond those image sets. The supplied abstract reports segmentation metrics and inference , not effects on diagnosis, monitoring decisions or patient outcomes.
Open question
Does the reported advantage persist beyond the 25 held-out images and 90 external images described in the abstract?
Source-stated
The authors propose a hybrid pipeline that uses an unmodified UNet for a coarse location prior and SLIC to recover lesion boundaries at full input resolution.
an unmodified UNet trained on 128 * 128 images provides a coarse location prior, while a training-free superpixel method (SLIC) recovers the boundary at the full input resolution.
Source-stated
The location prior automatically selects the superpixel count for each image, avoiding manual tuning.
The prior further drives automatic per-image selection of the superpixel count via an image-to-lesion ratio, eliminating manual tuning.
Source-stated
On 25 held-out 1024 * 1024 images, the proposed method achieved a mean Dice of 90.9%.
On 25 held-out 1024 * 1024 images, the proposed method achieved a mean Dice of 90.9% (IoU 87.9%)
Source-stated
On 90 external images from a second centre, the paper reports 76.8% for the proposed method and 69.3% for nnU-Net.
On 90 external images from a second centre, the advantage persisted (76.8% vs. 69.3% for nnU-Net).
Source-stated
The abstract reports that inference requires no GPU and takes 1.6 seconds per 1024 * 1024 image on a laptop.
Inference needs no GPU, requiring 1.6 s per 1024 * 1024 image
Escape Meridian interpretation
The supplied abstract reports image-set segmentation results but does not report effects on diagnosis, monitoring decisions or patient outcomes.
Accurate delineation at native resolution is challenging: CN typically occupy
Source and contributor context
Eshragh, Mohammadmahdi, Mohammed, Emad A., Far, Behrouz, Weis, Ezekiel, Shields, Carol L, Ferenczy, Sandor R, Crump, Trafford · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: cbc5b8b9-0acf-417b-876f-8592da2e5773.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T21:03:39.412Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: eb4825712b5b35ac3e37414ba3419f7c23fd6cc1d49c2fa8c42451af83865304. Source text hash: 6241951b9451f1a392607e7a574046dd001a32a087a8364c29a361b2ebea432f.
uncertain · 2026-11-08T21:03:39.412Z · anthropic/claude-opus-5.5
Publication: cbc5b8b9-0acf-417b-876f-8592da2e5773
attribution — pass: The draft attributes the pipeline, metrics and ablation findings to the paper's authors and the abstract, using phrasing such as 'the authors report' and 'according to the abstract.' Results are not presented as independently established.
Authors:Mohammadmahdi Eshragh, Emad A. Mohammed, Behrouz Far, Ezekiel Weis, Carol L Shields, Sandor R Ferenczy, Trafford Crump
summary — uncertain: Two problems. First, central terms are unexplained for a general reader: fundus, Dice, UNet, superpixel and nnU-Net (suggest term-library tooltips). Without them, 'mean Dice of 90.9%' has no meaning. Second, the headline in-distribution result appears without its comparison: the abstract reports 90.9% against nnU-Net 83.2%, Swin UNet 73.0% and Attention UNet 60.6%, all trained at the same resolution, with differences significant after Holm correction. Smallest repair: add those baselines and conditions, plus the stated difficulty that CN occupy under 2% of the image.
outperforming nnU-Net (83.2%), Swin UNet (73.0%), and Attention UNet (60.6%) trained at the same resolution, with all paired differences significant after Holm correction
date — pass: The originalPublished value 2025-09-29 matches the v1 submission date. It is not the later v4 revision (4 Nov 2026) or the feed date.
[Submitted on 29 Sep 2025 (v1), last revised 4 Nov 2026 (this version, v4)]
support — uncertain: The stated metrics, the no-GPU claim and the 1.6 s timing match the abstract. However, the labeled EM interpretation contrasts the method with 'training a segmentation model directly at that resolution' and carries no supporting excerpt. The abstract only says the baselines were 'trained at the same resolution,' without specifying which resolution, so this contrast is ambiguous. The excerpt cited for the outcomes-gap claim describes delineation difficulty, not the absence of clinical outcomes. The question and limitations are appropriately conditional.
trained at the same resolution
media — pass: The selected media is the verified direct link to the arXiv abstract page. It is the original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIME