U-Net: Convolutional Networks for Biomedical Image Segmentation

U-Net: Convolutional Networks for Biomedical Image Segmentation Olaf Ronneberger Philipp Fischer Thomas Brox

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

U-Net: Convolutional Networks for Biomedical Image Segmentation 2015 MICCAI / arXiv 1505.04597 ronneberger2015-unet ronneberger2015-unet.txt, the project's staged copy 6150cb065c2701e8b18886a0691747bb98e0ed61 arxiv-open medicine

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. biomedical image segmentation Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms overlap-tile prediction with mirrored extrapolation at the border: the missing context outside the image is manufactured by reflecting the image itself, so a tile can be segmented without the network ever seeing beyond the edge Quotation withheld because this record is not cleared for redistribution. the paper's claim that this teaches invariance WITHOUT annotating the transformations is made in the introduction, not in this span; tagged separately or not at all elastic deformation of the training images, generated from random displacement vectors on a coarse 3 by 3 grid, sampled from a Gaussian with 10 pixels standard deviation and interpolated to per-pixel displacements with bicubic interpolation Quotation withheld because this record is not cleared for redistribution. the size of that weight - w0 = 10 - is stated in a different sentence and is not claimed here a per-pixel loss weight, pre-computed from each ground truth segmentation by morphological operations, that both compensates class frequency and forces the network to learn the small separation borders introduced between touching cells Quotation withheld because this record is not cleared for redistribution. mechanisms a large number of feature channels in the expansive path, which is what carries context from the contracting path up into the high-resolution layers and makes the architecture symmetric Quotation withheld because this record is not cleared for redistribution. mechanisms batch size traded down to a single image so the input tile can be large, with the momentum raised to 0.99 so that the update still integrates many previously seen samples Quotation withheld because this record is not cleared for redistribution. mechanisms initial weights drawn so that each feature map starts at approximately unit variance, which is what stops one path through the network from saturating while another never contributes Quotation withheld because this record is not cleared for redistribution. Because convolutions are unpadded, border pixels are lost at every layer and feature maps must be cropped before concatenation. what this limits: every output map's extent relative to its input: THE CROPPING IS NECESSARY DUE TO THE LOSS OF BORDER PIXELS IN EVERY CONVOLUTION Quotation withheld because this record is not cleared for redistribution. Without the overlap-tile strategy the resolution the network can handle is bounded by GPU memory. what this limits: application to large images: THIS TILING STRATEGY IS IMPORTANT ... SINCE OTHERWISE THE RESOLUTION WOULD BE LIMITED BY THE GPU MEMORY Quotation withheld because this record is not cleared for redistribution. Part of what the network is trained to detect is inserted by the authors: the separation borders between touching cells are synthesised, not annotated. what this limits: a comparison of this network's boundary accuracy against a method trained on unmodified ground truth: the separation borders between touching cells are introduced by the authors and computed with morphological operations, so the two are trained on different targets Quotation withheld because this record is not cleared for redistribution. How heavily those borders count is set by two chosen constants in the loss. what this limits: the segmentation scores reported here: the pixel-wise loss weight that forces border learning is controlled by two values the authors chose, w0 = 10 and a sigma of about 5 pixels Quotation withheld because this record is not cleared for redistribution. The invariances the network has are the ones the augmentation supplied, and the paper says the corpus never contained them. what this limits: a claim that this network is robust to deformation: the invariance was taught by applying elastic deformations during training, and the paper states the annotated corpus did not need to contain those transformations at all Quotation withheld because this record is not cleared for redistribution. convolutional networks applied to biomedical image segmentation Quotation withheld because this record is not cleared for redistribution. evaluation on a public challenge test set whose segmentation maps are kept secret by the organisers eliminates: tuning against the test labels - the authors could not have done it even had they wanted to, which is a stronger guarantee than any promise a paper makes about itself · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. data augmentation data augmentation Quotation withheld because this record is not cleared for redistribution. elastic deformations elastic deformations Quotation withheld because this record is not cleared for redistribution. invariance invariance Quotation withheld because this record is not cleared for redistribution. weight map weight map Quotation withheld because this record is not cleared for redistribution. separation border separation border Quotation withheld because this record is not cleared for redistribution. the ground truth - PARTLY SYNTHESISED: the annotation is augmented with separation borders computed by morphological operations so that touching instances can be pulled apart, so the target is the annotation plus an artefact of the method Quotation withheld because this record is not cleared for redistribution. very few images - AS THIRTY, AND THEN AS PATCHES: the electron-microscopy training set is 30 images, and the paper notes that in patches the training data is much larger than the image count suggests Quotation withheld because this record is not cleared for redistribution. invariance - AS A PROPERTY OF THE AUGMENTATION: what the network is invariant to is what was applied to the training images, not what was observed in the microscopy Quotation withheld because this record is not cleared for redistribution. extends → sliding-window convolutional networks Quotation withheld because this record is not cleared for redistribution. contradicts → the assumption that deep networks need thousands of annotated samples Quotation withheld because this record is not cleared for redistribution. builds-on → the fully convolutional network of Long et al. [9] Quotation withheld because this record is not cleared for redistribution. extends → the fully convolutional network of Long et al. [9] Quotation withheld because this record is not cleared for redistribution. contradicts → the sliding-window patch strategy of Ciresan et al. [1] Quotation withheld because this record is not cleared for redistribution. elastic deformations —causes→ invariance elastic deformations invariance strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: 'This allows' = the augmentation by elastic deformations (previous sentence). Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · weight map —causes→ separation border weight map separation border EVIDENCE CORRECTED: RE-GROUNDED. The old span named only 'separation border'. This sentence names both and states the mechanism: the weight map is pre-computed IN ORDER TO force the network to learn the separation borders. — strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says the weight map forces the network to learn the separation borders; causes is loose. Quotation withheld because this record is not cleared for redistribution. 30 — training set size for the ISBI EM segmentation task Quotation withheld because this record is not cleared for redistribution. 7 — test-time averaging over rotated inputs Quotation withheld because this record is not cleared for redistribution. A contracting path captures context and a symmetric expanding path enables precise localization. about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. U-Net won the ISBI cell tracking challenge 2015 in these categories by a large margin. about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. annotated samples annotated samples Quotation withheld because this record is not cleared for redistribution. class label class label Quotation withheld because this record is not cleared for redistribution. data augmentation data augmentation about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. contracting path contracting path Quotation withheld because this record is not cleared for redistribution. expanding path expanding path Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://arxiv.org/abs/1505.04597.