CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning

CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning Pranav Rajpurkar Jeremy Irvin Andrew Y. Ng et al.

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.

CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning 2017 arXiv preprint 1711.05225 rajpurkar2017-chexnet rajpurkar2017-chexnet.txt, the project's staged copy a8d2157991e27e34585f78989e78a452b01ac7f8 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. pneumonia detection on chest radiographs 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 the binary cross entropy is weighted by the prevalence of the OPPOSITE class - w+ is the negative share and w- the positive share - so the rarer label carries the larger weight Quotation withheld because this record is not cleared for redistribution. mechanisms dense connections improve the flow of information and gradients through the network, which is what the paper credits for making the optimization of a very deep network tractable Quotation withheld because this record is not cleared for redistribution. mechanisms the network is not trained from scratch: the final layer is replaced and the rest starts from ImageNet weights Quotation withheld because this record is not cleared for redistribution. mechanisms the learning rate is cut only when the validation loss stops improving, and the returned model is the best validation checkpoint rather than the last Quotation withheld because this record is not cleared for redistribution. that this makes reader agreement the standard rather than the presence of disease is already recorded on this record as a limitation, not claimed here each radiologist and the model are scored against the OTHER FOUR radiologists' labels, taken as ground truth, and the reported figure is the mean of those four scores Quotation withheld because this record is not cleared for redistribution. mechanisms significance is taken from the DIFFERENCE between the two averages computed on the same bootstrap samples, and read off whether the 95% CI on that difference contains zero Quotation withheld because this record is not cleared for redistribution. mechanisms the heatmap is the weighted sum of the final convolutional feature maps, weighted by the final classification layer's own weights for that pathology, then upscaled and overlaid about this quote: 2026-08-16: FAILED GLYPH INSIDE THIS QUOTE. The staged text reads "Formally, Mc = (cid:88) k wc,kfk" - (cid:88) is the SUMMATION SIGN, so the formula is Mc = sum over k of w_c,k * f_k, the weighted sum of feature maps the sentence has just described in words. The prose on either side is intact and states the mechanism; only the symbol is lost. The quote is an exact substring of the staged file; the file is incomplete here, not the tag. Quotation withheld because this record is not cleared for redistribution. the paper identifies three limitations of its radiologist comparison what this limits: the radiologist comparison, which the authors themselves qualify: WE IDENTIFY THREE LIMITATIONS OF THIS COMPARISON Quotation withheld because this record is not cleared for redistribution. The paper announces three limitations of the radiologist comparison but enumerates only two (a 'First' and a 'Third'); the second is missing from the text. duplicate of: limitations[0] · what this limits: the completeness of that qualification: the paper announces three limitations and prints a FIRST and a THIRD with no SECOND. Verified in the version of record, arXiv 1711.05225v3 (25 Dec 2017), not only in our staged copy - so this is the paper's own text and not a lost sentence in extraction Quotation withheld because this record is not cleared for redistribution. Only frontal radiographs were shown to both model and radiologists, although a share of accurate diagnoses requires the lateral view. what this limits: the diagnostic setup for both model and radiologists: ONLY FRONTAL RADIOGRAPHS WERE PRESENTED ... BUT IT HAS BEEN SHOWN THAT UP TO 15% OF ACCURATE DIAGNOSES REQUIRE THE LATERAL VIEW Quotation withheld because this record is not cleared for redistribution. Neither model nor radiologists could use patient history, which is known to change radiologist diagnostic performance. what this limits: the comparison against clinical practice: NEITHER THE MODEL NOR THE RADIOLOGISTS WERE ... PERMITTED TO USE PATIENT HISTORY, WHICH HAS BEEN SHOWN TO DECREASE RADIOLOGIST DIAGNOSTIC PERFORMANCE Quotation withheld because this record is not cleared for redistribution. The standard the model is measured against is the other radiologists' labels, not the presence of disease. what this limits: the claim that this model exceeds practicing radiologists: the F1 score for each radiologist and for the model is computed against each of the other four radiologists' labels as ground truth, so the comparison is agreement with those four rather than detection of disease Quotation withheld because this record is not cleared for redistribution. The training labels were produced by text extraction from reports rather than by anyone looking at the images. what this limits: the labels this model was trained on: they were assigned by automatic extraction methods applied to radiology reports, so a label records what a report said rather than what an image shows Quotation withheld because this record is not cleared for redistribution. Every image the extractor did not mark becomes a negative example. what this limits: the negative class in this task: images are negative when pneumonia is not among the annotated pathologies, so absence of a label is used in place of absence of the condition Quotation withheld because this record is not cleared for redistribution. The model reports fourteen findings while the headline number describes one of them. what this limits: the performance figure quoted for this model: the network outputs a vector of binary labels over fourteen pathology classes, while the radiologist comparison concerns the pneumonia task alone Quotation withheld because this record is not cleared for redistribution. a deep convolutional network (DenseNet) applied to detect disease from chest X-rays at radiologist level Quotation withheld because this record is not cleared for redistribution. four practising academic radiologists annotating the same test set, used as the comparison for the model eliminates: a number that looks good against no reference - the baseline here is people doing the task on the same images · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. F1 score F1 score Quotation withheld because this record is not cleared for redistribution. ground truth ground truth Quotation withheld because this record is not cleared for redistribution. pneumonia detection task pneumonia detection task Quotation withheld because this record is not cleared for redistribution. pathology labels pathology labels Quotation withheld because this record is not cleared for redistribution. automatic extraction methods automatic extraction methods Quotation withheld because this record is not cleared for redistribution. 'exceeding practicing radiologists' - AS AGREEMENT WITH FOUR OF THEM: the F1 for the model and for each radiologist is computed against the other four radiologists' labels, so the claim is about concordance with a particular panel and not about detecting pneumonia in a patient Quotation withheld because this record is not cleared for redistribution. a pathology label - AS A FACT ABOUT A REPORT, NOT AN IMAGE: the fourteen labels are assigned by automatic extraction from radiology reports, so the target the network learns is what a radiologist wrote down, filtered through a text-mining tool Quotation withheld because this record is not cleared for redistribution. 'negative' - AS NOT-LABELLED: every image without pneumonia among its annotated pathologies becomes a negative example, so the class means the extractor found nothing rather than that nothing was there Quotation withheld because this record is not cleared for redistribution. builds-on → convolutional neural networks (deep learning) Quotation withheld because this record is not cleared for redistribution. builds-on → DenseNet (Huang et al., 2016) Quotation withheld because this record is not cleared for redistribution. builds-on → the ChestX-ray14 dataset of Wang et al. (2017) Quotation withheld because this record is not cleared for redistribution. builds-on → the evaluation protocol of Wang et al. (2017) and Yao et al. (2017) Quotation withheld because this record is not cleared for redistribution. pathology labels —operationalised-as→ automatic extraction methods pathology labels automatic extraction methods strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says labels were annotated USING automatic extraction methods; operationalised-as is loose. Quotation withheld because this record is not cleared for redistribution. 121 — depth of the DenseNet CNN (CheXNet) Quotation withheld because this record is not cleared for redistribution. 112,120 — chest X-ray images in the ChestX-ray14 dataset 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. 14 — thoracic disease labels detected 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. 420 — size of the radiologist-annotated test set Quotation withheld because this record is not cleared for redistribution. > 0.05 — margin over previous state of the art (emphysema) Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED - THE POINTER WAS WRONG, NOT THE CLAIM) · CheXNet is a 121-layer CNN, trained on the ChestX-ray14 dataset of over 100,000 frontal-view X-rays covering 14 diseases. EVIDENCE CORRECTED: RE-GROUNDED - THE POINTER WAS WRONG, NOT THE CLAIM. The old evidence was verbatim, anchored and correctly cut, and it contained NONE of this claim's numbers (121, 14). The paper states the claim in one sentence elsewhere and the tagger did not use it. the project, 2026-08-23: 'WE HAVE SPENT THE DAY VERIFYING THAT QUOTES ARE REAL AND NONE OF IT ASKS WHETHER THEY ARE RELEVANT.' Confirmed by reading the claim, the old span and the candidate together; the candidate carries every number AND the claim's own content words. — 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. Extended to all 14 thoracic diseases, CheXNet achieves state-of-the-art results on every one. 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. chest X-ray image chest X-ray image Quotation withheld because this record is not cleared for redistribution. F1 metric F1 metric 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. CheXNet CheXNet Quotation withheld because this record is not cleared for redistribution. ChestX-ray14 dataset ChestX-ray14 dataset about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. bootstrap bootstrap 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. confidence-interval confidence-interval 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. mean mean 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. standard-deviation standard-deviation Quotation withheld because this record is not cleared for redistribution. statistical-significance statistical-significance 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.

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/1711.05225.