Every change to the list and to this site, newest first. The list does not tag releases: it is refined continuously, so changes are recorded by date rather than by version. To cite a fixed snapshot, use a commit hash.
This page is generated from CHANGELOG.md in the repository on every build. 22 entries, from 2026-10-10 to 2026-10-16.
— Dataset and benchmark pages, and three guides
Added
/datasets/<slug>/ and /benchmarks/<slug>/ (English and Chinese) for the 21 datasets and 11 benchmarks in the directory with real content: a description, an official link, and at least one topic page or paper in the directory that uses them. Each page shows the data the entry ships (projection and/or image, as its note states), modality, tasks, the dataset's own licence, official links and DOI where the record has one, the paper that describes it, the papers in the directory that name or link it, its topic pages, related entries and guides. Each has its own title, description, canonical URL, hreflang alternates, a BreadcrumbList and a schema.org Dataset block built from the same record as the visible page. The other 28 entries, mostly segmentation and detection collections with no reconstruction topic or paper in the directory yet, stay in the directory tables; build_site.py prints each one it skips and why. Which entries get a page, their URL slugs and the aliases used to find the papers that use them live in the new data/resources.yaml.
/datasets/, an index of every dataset and benchmark page.
Three guides under /guides/, with a /guides/ index: sparse-view CT reconstruction, low-dose CT denoising and how to choose a CT reconstruction dataset. Each opens with the answer, compares the options in tables generated from the directory (method families, datasets, benchmarks, toolkits, start-here papers), shows a last-updated date and ends with a short FAQ whose FAQPage JSON-LD is built from the visible text. Prose lives in data/guides.yaml; every count in it is filled from the live data.
Topic pages link the dataset and benchmark pages they list (the official site stays one click away) and the guides that cover them; the About page links /datasets/, /guides/ and each guide. All new pages are in sitemap.xml and llms.txt. The homepage is unchanged.
build_site.py fails if a new page lacks its canonical, hreflang, breadcrumb or JSON-LD, is missing from the sitemap, has no inbound link, or if a guide links a page that is not generated. tests/test_seo.js also checks that each Dataset block matches the visible name, description, licence, DOI and citation, and that each guide's FAQ and last-updated date match its JSON-LD.
Fixed
A paper title that starts with its abbreviation followed by a comma (LoDoPaB-CT) no longer renders as "LoDoPaB-CT: , a benchmark …" on topic pages.
— A changelog page, and the homepage described as a data catalog
Added
/changelog/ and /zh/changelog/, generated from CHANGELOG.md on every build by scripts/changelog.py: every dated entry, newest first (entries that share a date keep their order in the file). The Chinese edition shows each entry in Chinese, read from CHANGELOG.zh-CN.md and matched to its English original by date and title; an entry with no translation yet is shown in English, marked lang="en", and the page says so. Each edition has its own title, description, canonical URL, hreflang alternates and a BreadcrumbList; both are in sitemap.xml and llms.txt, and are linked from the About page and the footer. The build fails on an undated ## heading in CHANGELOG.md, or on a date more than a day ahead of the build.
A DataCatalog block in the homepage JSON-LD (English and Chinese), beside the existing Organization, WebSite and ItemList. It names the three tables the homepage shows — papers, datasets, benchmarks — with counts filled from the live data, dateModified from last_updated in data/meta.yaml, and the Organization as creator and publisher. build_site.py refuses to write a catalog whose counts or date disagree with data.json.
tests/test_seo.js checks both, with no packages, and runs in validate.yml.
— Topic pages for every topic, and an FAQ on the About page
Merged as #2 (squash 3459829).
Added
/topics/<id>/ and /zh/topics/<id>/ — a hub page for each of the 16 topics in data/meta.yaml (all but other), in English and Chinese. Each carries a hand-written introduction, a short "start here" reading list chosen by the maintainers, the topic's 30 newest papers, the related datasets, benchmarks and toolkits, and links to the sibling topics. The editorial text lives in the new data/topics.yaml; the paper table, counts and related resources are generated, so they cannot go stale. Every page has its own canonical URL, hreflang alternates (en, zh-Hans, x-default), a BreadcrumbList and a CollectionPage block, and is in sitemap.xml and llms.txt.
About page (English and Chinese): "Browse by topic" and a six-question FAQ — which datasets suit sparse-view CT, which toolkits are used for low-dose CT, how entries are verified, how often the list is updated, how to cite it, and how to suggest a paper or report a mistake. The FAQPage JSON-LD is generated from the visible FAQ text, so the two cannot disagree.
build_site.py checks that every start id and also name in data/topics.yaml exists, that each topic page carries its canonical, hreflang, JSON-LD and sitemap entry, and that the About page links every topic page.
— Ownership verified with Google Search Console and Bing Webmaster Tools
Added
google_site_verification and bing_site_verification in site.config.yml. build_site.py emits them as the google-site-verification and msvalidate.01 meta tags in the homepage <head> (English and Chinese). Both services re-check the tag, so removing either un-verifies the site; the config comments say so.
— llms.txt, and IndexNow pings after each deploy
Merged as #1 (squash 6736652). No change to existing pages, styles or data.
Added
/llms.txt, a plain-Markdown map of the site for AI answer engines (the llmstxt.org proposal): what the list covers, links to both editions, the About and Contact pages and both data.json exports, and every topic with its paper count. Counts and topics come from the same payload as the page, so it cannot go stale. Cheap to ship; whether anything reads it is unproven.
IndexNow.indexnow_key in site.config.yml; the build writes the key file site/<key>.txt at the site root, and deploy.yml gained a job that, after each deploy, posts every URL in sitemap.xml to api.indexnow.org (Bing, Yandex and the other IndexNow engines). It is best effort — a failed ping never fails the deploy. Google does not use IndexNow; it reads the sitemap submitted in Search Console.
— Clean-up after the blurb work
Removed
TODO-summaries.md — the checklist of papers still lacking a blurb. Coverage reached 401/401 in both languages, so the file only invited work that no longer existed.
Fixed
tests/test_site_dom.js asserted that at least one paper had no blurb, to prove such a title stays inert. That premise expired the moment coverage hit 100%, quietly turning the assertion into a permanent false failure. It now checks what actually has to hold — has-blurb and data-blurb agree, so a title without a blurb can never become hoverable.
— README is a landing page; the full lists live on the website
Changed
The READMEs keep their structure — header, stats strip, Papers, Toolkits, Datasets, Benchmarks, Learning, Contact, License — but no longer expand the entries one by one. The Papers section is a topic grid whose cells link to the website with that topic already selected; the other four sections each link to their site tab. Entry notes and the dataset legend still render on the site, which is the full listing.
scripts/build.py no longer contains the per-entry renderers (render_papers, render_toolkits, render_datasets, render_benchmarks, render_learning); the i18n keys only those renderers used are removed.
— Stats show the last-updated date, not the paper-year span
Changed
The README stats strip, the site stats row and the coverage sentences now show "Updated 2026-10-15" (from the new last_updated field in data/meta.yaml) instead of the paper-year span "1917–2026", so a visitor can tell at a glance whether the list is still maintained. Bump last_updated whenever content changes; scripts/build.py --check fails if it is missing or not a YYYY-MM-DD date.
— Bilingual hover blurbs, and corrections from an outside review
Added
A paper title explains itself on hover. The title used to be an <a> to the same URL as the paper link beside it, so clicking either went to the same place. It is plain text now, and papers we hold a blurb for show one in a card on hover, tap or keyboard focus. No second page — the links below are the way out. Papers without a blurb stay inert, and a title= attribute keeps a fallback for readers with JavaScript off.
data/summaries.yaml — one-paragraph blurbs for 256 of 399 papers, each built from verbatim sentences of the official abstract (the condense step chooses sentences, it never rewrites), so a blurb can be incomplete but never invented. Fetched in two passes: arXiv by id (158/158), Crossref by DOI (47 of 237 — most publishers hand Crossref no abstract), then arXiv by title behind a ≥0.90 similarity gate (+51).
scripts/oneoff/fetch_summaries.py and fetch_summaries_by_title.py, both rate-limited and cache-backed (--offline re-derives from the cache).
The blurbs are bilingual.data/summaries.zh.yaml carries the Chinese wording for all 256, so the Chinese site shows Chinese and the English site shows English. They live in their own file rather than inside data/i18n/zh.yaml, which is a curated overlay that a few hundred machine-length entries would swamp; scripts/export_site_data.py picks the file matching the language being built and falls back to English for anything untranslated, so a gap shows up as visible English rather than an empty card.
scripts/oneoff/translate_summaries.py — applies a batch of translations and reports what is still missing (--status).
TODO-summaries.md — the 143 papers still without a blurb, with a short how-to. Nothing is wrong with them; they simply show no card.
Fixed
An outside review of the site turned up a batch of real errors. Each was re-checked against a primary source before being applied:
Eight papers filed as preprints had already been published: CTO (ECCV 2026), NAB-GS (MICCAI 2026), PMF-STGR (PMB 70(16):165011), DREME-adapt (Med Phys 52(11):e70119), the accelerated-INR paper (ISBI 2025), EAGLE (J. Med. Imaging 12(1):014001), DDDM (IEEE SPL 31:1279–1283) and UncertaINR (TMLR 2023). Preprints 67 → 59. Note that EAGLE's arXiv page still says "Preprint" — the journal page is what settles it.
Five title typos: "Denosing" → "Denoising", "Low-Dose Ct" → "Low-Dose CT", "HorusEye: a …" → "HorusEye: A …", "multi-domain … swin transformer" → "Multi-domain … Swin Transformer", "for Cone beam CT" → "for cone-beam CT".
Twenty-four papers whose title already begins "ABBR: …" had no abbr, so the site showed no leading tag and the column looked ragged. Added — which also makes those abbreviations searchable.
One blurb still carried the raw <p> markup from a publisher's Crossref record; unescaped and stripped.
Not changed, deliberately
The TomoPhantom and LoDoPaB-CT entries the review called duplicates are intentional: each cross-references the other, and the LoDoPaB one is the citation for the dataset rather than a second copy of it.
The review's claim that the Chinese and English sites had drifted was not reproducible — both served 399 papers and the same build date.
— Seven papers and one toolkit from an arXiv sweep
Scanned arXiv across computed tomography, sparse-view, limited-angle, CT reconstruction, cone-beam and low-dose CT, newest first, and took the work that belongs on this list.
Added
PhyDiCT (MICCAI 2026) — training-free plug-and-play CT reconstruction from a few X-rays: a differentiable Beer–Lambert forward model steers a text-conditioned diffusion prior through Split Gibbs sampling, jointly optimising projection fidelity and image prior. Sparse-view, with code.
OX-NeRF — X-ray tomography from ultra-sparse views (10 or fewer) by pairing cross-scene feature learning with per-scene multi-resolution hash grids.
Gradient-based trajectory optimisation for sparse-view CBCT — every source pose becomes a continuous variable moved jointly by gradient ascent on the scanner's kinematic manifold, so the scan path is designed rather than picked from a fixed candidate pool.
Gaussian material fields — multi-energy CT material decomposition from shared anisotropic 3D Gaussian primitives with independent per-material coefficients.
LINGO — latent initialisation and gradient optimisation for sparse-view X-ray novel-view synthesis and CT reconstruction with 3D Gaussian splatting.
Model-based iterative reconstruction with view-dependent detector displacements — folding measured detector motion into the CBCT forward model instead of treating the geometry as nominal.
Deep Pseudo-Proximal Map — a self-supervised data-fitting agent standing in for the proximal map inside an iterative reconstruction loop.
CTrex (toolkit) — a research-oriented GPU framework that exposes the projection–correction–backprojection pipeline down to individual CUDA kernels, so non-ideal acquisition effects can be written into the operators rather than bolted on as corrections; geometry rides on per-projection homogeneous view matrices, keeping kernels agnostic to circular, helical, offset and conveyor-belt trajectories. Listed through its preprint: the paper announces no code repository, and the entry will point at a repo once one appears.
Fixed
Geometry-Aware Diffusion Approximate Posterior Sampling was filed under limited-angle only, though both its title and its method cover sparse-view CT as well. Added sparse-view — the same kind of under-tagging that left DC-PnPDP unsearchable by abbreviation the day before.
Deliberately left out
Scanned and rejected: CPF-DDNM (a general diffusion posterior sampler whose abstract never mentions CT), randomised batch-sampling coordinate descent (pure optimisation theory), EviDent-CBCT (report generation from dental CBCT, not reconstruction), a deformable CT–US registration method, and the current crop of 3D Gaussian-splatting and NeRF papers for general scenes — those are computer-vision view synthesis, not tomography.
— DC-PnPDP findable by its name
Fixed
DC-PnPDP (ICML 2026) was already listed, but only under its full title, so a search for "DC-PnPDP" missed it. It now carries the abbreviation, its project page, and the sparse-view and limited-angle topics its CT results cover (SVCT-20, LACT-90).
— Cloudflare Web Analytics, and the privacy page that says so
Changed
analytics_token in site.config.yml is now set, so the list page (English and Chinese) carries the Cloudflare Web Analytics beacon. The domain is not proxied through Cloudflare, which is why the zone dashboard showed zero visitors; the JS beacon counts visits regardless. It sets no cookies and no local storage.
Privacy page (English and Chinese, plus the maintainer copy in web/pages/privacy.html): the "Analytics: none" section now describes the beacon, what Cloudflare receives, and that only aggregate numbers are seen. Updated 12 October 2026, as the page promised it would be before any change.
— Search, sorting and language-switch fixes
Fixed
An arXiv id typed into the search box no longer jumps to the wrong page. Toolkit, dataset, benchmark and learning records indexed their whole URL, so every number inside it became a searchable token and an id matched an unrelated record. They now index the host only, which is what the code always said it did. The bare arXiv id is indexed separately — 151 of the papers link to one — so searching 2607.14415 finds $K$-NeAS, the paper it belongs to, instead of a dataset whose URL happens to contain 2607.
Enter needs a real hit before it navigates. With no paper match, Enter opened the top-ranked result however weak it was. It now opens only at a score of 2 or more, which is what a name or metadata hit earns; a lone weak body hit (~0.7) filters the list instead, so a stray keystroke cannot send someone off-site.
Title sort no longer files $K$-NeAS first. The sort used the raw title, and $ sorts ahead of every digit and letter; it now sorts on the title as displayed, with the inline TeX delimiters stripped. The comparator also returns 0 for equal keys, which a -1/1-only comparator never did.
The year filter's "and later" label is translated. It was built as a hardcoded English string; it now reads the existing yearAndLater key, which the Chinese overlay already provides.
Switching language keeps the filters. The switch links are written at build time and carry no query string, so /zh/?tab=datasets&q=… lost every filter on switching. web/lang.js appends the current query once at load; both templates load it.
Docs
SITE.md still described the search as "filters as you type … multiple words are AND-ed", which stopped being true when ranking replaced the strict AND. It now describes the dropdown-then-Enter behaviour and why Enter has a threshold.
— Corrected two challenge descriptions and two dataset records
Fixed
AAPM DL-Sparse-View CT Challenge (and its dataset entry): the challenge is the 2021 one, not 2022, and it asks for recovery of simulated 2D breast-CT phantoms from ideal noiseless 128-view fan-beam data (4,000 training pairs, 100 test cases). The earlier text described a 2/4/8/16/32/64-view re-projection of the 2016 clinical cohort, which is not what the challenge provides. Dataset modality and licence now follow the AAPM page and the Zenodo record (CC BY 4.0).
Helsinki Tomography Challenge 2022: the winner was a purely learned network trained on synthetic data; the only model-based entry (CIL) placed third. The note had this the wrong way round.
LDCT-and-Projection-data licence: TCIA version 7 opened the chest and abdomen scans under CC BY 4.0; the head scans stay under NIH controlled access. It was recorded as CC BY-NC 4.0.
— Three CT foundation models, and the CT-RATE dataset
Added
TAMP, DeepSparse and LCTfound. Each was resolved against Crossref before being written down — title match 1.000 for all three — rather than taken from the PDF it was harvested from, because the markdown those PDFs produce carries the arXiv ids of the references, not of the paper itself. TAMP (Nat Commun 2026) covers low-dose, sparse-view and limited-angle in one backbone; DeepSparse (TMI 2026) is the sparse-view CBCT one; LCTfound (Nat Commun 2025) spans eight tasks of which CT enhancement and sparse-view reconstruction are the reconstruction-relevant ones.
CT-RATE, in the image-domain half of the datasets: 50,188 reconstructed 3D chest CT volumes with a paired radiology report each. It gets no projection marker — there are no raw sinograms — which is worth knowing before reaching for it to test a forward model. Its landing page does not state a licence, so it is recorded as see source rather than guessed at.
scripts/oneoff/verify_foundation_papers.py and its output, so the batch's provenance sits in the repository beside the entries it produced.
Fixed
HorusEye had no code link. Its repository is now recorded.
Deliberately left out
Three papers came out of the same batch and were not added: CT-RATE's own paper (CT-CLIP / CT-CHAT), Merlin and TAP-CT. All three are CT foundation models, but what they contribute is vision-language retrieval, report generation and frozen-feature classification — not reconstruction. The bar in MAINTENANCE.md is a paper on CT reconstruction, and a reader who came here for reconstruction methods would find them noise. Recorded here rather than quietly dropped.
— The contributor graph is cslht11-only again
Fixed (attribution)
A merged Dependabot pull request put dependabot[bot] in the contributor graph. Squash-merging attributes the new commit to the pull request's author, so main picked up a commit authored by dependabot[bot] — ranked second, with one commit, next to cslht11's fourteen. The commit is rewritten as a cslht11 commit; the change it carried (the PyYAML floor) is kept exactly as it was.
.github/dependabot.yml is deleted. A refs/pull/N/head ref cannot be removed — the API answers refs/pull/* is read-only (HTTP 422) — so from the moment Dependabot opens a pull request its author is in that graph for good, whether the pull request is merged, closed or left open. There is no setting that scopes the graph to one author. Version bumps are applied by hand as cslht11 commits from here on. Dependabot security alerts are a separate setting that needs no configuration file, and stay on.
Added (authorship guards)
hooks/ and scripts/install-hooks.sh. The two guards that keep this from happening again lived only in one clone's .git/hooks/, which is to say they were one recloning away from not existing. Both are now in the repository and installed by one command, which the repository map now says to run after a clone.
hooks/pre-commit — refuses a commit whose user.email is not cslht11's.
hooks/pre-push — refuses any push containing a commit not authored by cslht11. This is the one that would have stopped dependabot[bot]: squash-merging signs with the pull request author's identity, which local git config never sees, so no pre-commit hook can catch it.
— Licences made precise, structure written down, payload 16% smaller
Fixed (licensing)
The content/code boundary was wrong in two places.LICENSE-content listed site/ as content, but the build copies web/*.js and web/*.css into it verbatim — so the same bytes were claimed under both licences. It also missed data/i18n/*.yaml entirely (data/*.yaml does not match a subdirectory), which is where all the translated prose lives. Both files now list their paths explicitly, including the site/*.js and site/*.css carve-out.
MAINTENANCE.md explains why the code is not under the same licence as the content — CC advises against its licences for software: they say nothing about patents or about source versus object form, NonCommercial is not an open-source licence (the OSI and the DFSG both reject it), and ShareAlike is viral over code. The consequence of NC is stated plainly too, rather than left implied.
Added (structure)
A repository map in MAINTENANCE.md — every directory and file, plus a "to change X, edit Y, then run Z" table. Written for someone arriving with no context, which is what the file is for.
tests/, in the repository. The search tests lived in a private ops directory; a test that is not next to the code it tests is a test that rots. tests/test_search.js needs no packages and now runs in validate.yml after the site build; tests/test_site_dom.js drives the real pages in jsdom and stays a local pre-push check on purpose — a workflow that can go red because a registry was unreachable teaches people to ignore red, which is the argument SITE.md already makes for keeping the deploy workflow off pull requests.
Changed (maintenance)
GitHub Actions brought up to current majors: checkout v4 → v7, setup-python v5 → v7, upload-pages-artifact v3 → v5, deploy-pages v4 → v5. The deploy log had been warning that deploy-pages@v4 still targets Node 20.
requirements.txt pins PyYAML to >=6.0.3,<7. The floor is the first 6.x with the patched loader; the ceiling keeps a major bump a deliberate change with the tests run, not something CI picks up overnight.
Changed (payload)
The _q search blob is gone (payload schema 1 → 2). It was a lower-cased concatenation of title, abbreviation, venue, topic names, note and link hosts, shipped so the client could substring-match one string. web/search.js builds a weighted index in the browser from the structured fields instead, so the blob was duplicated content with no remaining reader — ~68,000 characters of it. index.html 541 KB → 472 KB, data.json 263 KB → 192 KB, the whole site 868 KB → 727 KB, and one less derived field to keep in sync.
— The site's search actually works now, and stops throwing
Two name collisions were breaking the site's JavaScript. Both were pre-existing, both were invisible in the served HTML (which is server-rendered), and together they explain why the search box did nothing.
renderPapers declared var t = titleParts(p), shadowing the i18n helper t() used a few lines below it. The resulting TypeError fired on every render, which aborted init() before it wired the search box — so typing in the filter never did anything.
renderToolkits used function (t) as its map callback, the same collision, and it threw for the 14 toolkits that carry a homepage.
The top-bar box declared var t for its debounce timer, shadowing its own t(): the first keystroke turned t into a timer id, and the next render threw. The dropdown now survives more than one keypress.
debounce and the /-to-focus handler had the same shadowing shape without calling t(), so they worked by luck; both were renamed while fixing the others.
Added: a real search core (web/search.js)
The old search was a case-insensitive substring test over one flat blob, with strict AND semantics and no ranking — so a query had to be an exact substring, one extra word emptied the list, and results kept whatever order the data was in.
CJK tokenising. Chinese is not space-separated, so a run of ideographs is indexed as both its characters and its adjacent pairs: 「稀疏视角」 becomes 稀 疏 视 角 + 稀疏 疏视视角. 「稀疏重建」 now finds a record titled 「稀疏视角 … 重建」 even though the user's two words never sit next to each other.
Field weighting. One weighted token map per record (title 5, abbreviation 4, metadata 2, everything else 1), keeping the highest weight per token, so a title hit outranks the same word buried in a note.
Ranking. Token weights scaled by query coverage, plus a bonus for the query appearing verbatim. Records matching nothing are dropped, but a partial match is no longer thrown away: ranking is a stable sort applied after the user's own sort, so ties keep the order they chose.
Substring fallback for Latin tokens. Exact token equality would lose "mamba" → DenoMamba and "splat" → Splatting, which the old substring search did find, so a query token with no exact key falls back to index keys containing it, at a discount.
Highlighting everywhere it matters — results, the dropdown, and the topic line of each paper row (a Chinese query usually matches the topic, not the English title, so without that the results would look like nothing had matched).
The top-bar box searches all five sections, not just papers, and Enter now filters the paper list only when the query actually found papers; otherwise it opens the best match.
Verified by two test suites (54 assertions on the core, 23 driving the built pages in jsdom), including that a render throws nowhere — the check that would have caught the collisions above.
Changed
No web font, and said so. The system UI face plus the platform's own CJK family is faster and better-looking here, and it keeps the page free of third-party requests and of any font-licence question. The stack gained system-ui and a fuller CJK fallback chain so macOS, Windows and Linux agree; the decision is documented where it is made.
The LoDoPaB-CT paper title no longer repeats "(Scientific Data)" — its venue already says it.
— Three external audits worked through, and the contributor list cleaned up
Three independent audits of the repository — data, i18n, links, CI, docs — were checked finding by finding rather than applied wholesale. Every repair was re-derived from Crossref, the publisher's record or the project's own primary source before it was made; several audit claims turned out to be wrong and were dropped or narrowed (see below).
Fixed (paper and dataset records)
Four wrong years or venues, each confirmed against the record at the DOI: LoDoPaB-CT 2017 → 2021 (Scientific Data), MAP-NN 2021 → 2019 (Nat Mach Intell), DL-PICCS 2019 → 2021 (Med Phys), DREAM-Net TMI 2022 → JBHI 2023 — a mistake the DOI's own prefix, 10.1109/jbhi..., already contradicted. DenoMamba 2025 → 2026, following the volume/issue convention the other IEEE entries use.
LoDoPaB-CT's title carried leftover markdown ((**Scientific Data**) -->) from a link removed by hand, and FSTensor's title was missing its separator ("FSTensorFull-spectrum…", now "FSTensor: Full-spectrum…").
Two duplicate pairs removed, keeping the peer-reviewed record in each case: CasRedSCAN's 2020 preprint against its TMI 2021 paper, and the CAIP 2025 registration paper against its earlier workshop-titled arXiv posting. Papers 391 → 389.
Dataset facts corrected against the primary sources. LDCT-and-Projection-data claimed "four dose levels (full / quarter / 10% / 5%)"; the challenge released routine dose plus one simulated reduction, and TCIA states 25% for head/abdomen and 10% for chest. VerSe claimed "3,741 CT/CBCT spine scans" — it is 374 CT scans from 355 patients, with no CBCT. Medical Segmentation Decathlon has six CT tasks, not four. KiTS23 has 599 scans (489 train + 110 test), not 489 across "three editions". RibFrac's two numbers were swapped: 660 scans carrying ~5,000 annotations. STOIC 2021 is 10,735 French first-wave scans, not "23,000+ multi-centre". LUNGx is a diagnostic classification task, not nodule detection. CHAOS's CT is portal-venous contrast, not low-dose. AutoPET ran four editions by 2025, not three.
sigpy ships no Radon primitives and RadIO is not a Radon/FBP package — both notes now describe what the code actually contains. LEAP's expansion was invented (it is LivermorE AI Projector for Computed Tomography). libcbct is CC BY-NC-SA 4.0, which GitHub's API reports as "Other". CBCT Scatter Correction Tool has no source in its repository — it ships a Windows executable — so it no longer claims a language.
Two splat-based CBCT metal-artifact papers were kept, not merged. Both exist and both are peer-reviewed (CVPR 2026 and Computer Graphics Forum 45(2) 2026, same group); each now carries a note saying so, so the near-identical titles read as a pair rather than as a duplication error.
generative-modeling-in-sinogram-domain now links its published TRPMS DOI (10.1109/trpms.2023.3309474) instead of the arXiv posting, and two Transformer papers and two purely-learned papers lost topic tags that did not describe them.
Chinese strings: the typo 高斯泼澅 fixed and the three competing renderings of "Gaussian splatting" unified.
Fixed (build and rendering)
paper_head() treated any title that merely started with the abbreviation as a prefix match, so an abbreviation like "A" would have mangled "A benchmark …". It now requires the abbreviation to be the first token, and drops an abbreviation repeated in trailing parentheses. web/app.js mirrors both rules and the two implementations cross-reference each other so they cannot drift.
The dataset marker stripper only recognised the English full stop, so the Chinese ToothFairy2 note — which had translated the marker itself — printed the sentence in both the Data and the Description column. The marker stays English now, as everywhere else.
A row can be renamed by an overlay, not only have its note translated. That is what learning."CT Principles and Algorithms".name was for, and the key had been silently dead; the site payload keeps the English name as its stable key and carries the translation as label.
tableDescription sat in the readme: section but is read through ov.ui(), so it could never be found and the Chinese README printed the English header "Description". Moved into ui: in both language files, together with ten readme: keys that reach neither README nor site — two of which were actively misleading: licenseBody1 stated CC BY 4.0 with commercial reuse (the licence is CC BY-NC-SA 4.0) and contributing* linked to CONTRIBUTING.md, deleted on 2026-10-10.
The README's site URL is no longer hardcoded.build.py reads the origin from site.config.yml, the single place the docs promise it lives, so a domain change can no longer leave the README badge pointing at the old address.
new entry label created. The issue form referenced it, but it did not exist — GitHub drops an unknown label silently, so suggestions arrived unlabelled.
Changed
Journal venue strings unified: the IEEE prefix dropped from TNS, TIT, TSP, SPM and JSAIT to match the bare abbreviations used everywhere else (IEEE Access keeps it, being the journal's actual name), and MedIA → Med Image Anal.
Chinese terminology unified (同行评审, 工具包, 出版方), plus wording corrections: "on-board imaging" is 机载成像 rather than 在轨成像, "beam hardening" is 射束硬化 throughout, and topics.diffusion is described as covering generative priors generally — which its own name ("Diffusion & Generative Priors") already claimed.
Three Chinese-language books and the XDesign entry now link or describe what the source actually is: two books gained their publisher's page, the 1992 one is documented as having no publisher page (so it links a title search), the Kak & Slaney note says "chapter PDFs" rather than "Free PDF", and XDesign is described as the phantom and acquisition-simulation package its README documents, not as "Differentiable".
validate.yml gained permissions: contents: read; check_links.py's disabled certificate check is now documented and justified rather than silent.
Corrected SITE.md's description of what CI does: the README-freshness guard lives in validate.yml and runs on pull requests too, while deploy.yml runs only on pushes to main. The two same-day 2026-10-10 entries were given distinct titles.
Not changed, deliberately
The workshop venues ("SC Workshops", "NeurIPS Workshop", "Off-Grid Workshop") are three different workshops, not three spellings of one; forcing a single string would lose information.
The 17 truncated ids (e.g. …-for-low-) stay as they are: id is the permanent /p/<id> key, so the generator was fixed instead of the data.
oneoff/'s 2.2 MB of harvested JSON stays: it is the provenance record that lets every entry be traced back to how it was found, and check_links.py still writes into it.
— 102 more papers, the whole list in Chinese, and a new home
The largest single addition so far. The list moves to a repository of its own and becomes bilingual end to end.
Content
391 papers (was 289), spanning 1917–2026 (was 1976) across 17 topics, plus 44 toolkits (was 43), 40 datasets (was 34), 19 benchmarks/challenges and 29 learning resources. 328 of 391 (84%) carry a peer-reviewed venue.
Radon 1917 added: the paper that introduces the transform every analytic method here inverts, published 54 years before the first CT scanner. No DOI exists for the original, so the entry links P. C. Parks's English translation in IEEE TMI (1986) and records the original Leipzig imprint as its venue.
57 papers added from a Web of Science export, kept only where Crossref confirmed a journal article and the work is a reconstruction method rather than an application that merely uses CT. A further 44 came from cross-checking a review manuscript's bibliography.
Language
The list, the site chrome, the static pages and both READMEs now exist in English and Chinese. Interface strings live in data/i18n/; English is the source of truth and every other language is an overlay that falls back to it.
/zh/ serves the Chinese site, with its own canonical URL, hreflang pair and sitemap entries.
Site
description is now filled in at build time. It used to spell out "289 papers", which went stale the moment a paper was added — and this is the text search engines show.
Fixed the light/dark toggle: it had one listener in app.js and another in chrome.js, so a click switched the theme twice. It now lives in one inline script in <head>, applies the saved choice before first paint, and the button shows the icon for the theme it will switch to.
The EN / 中文 switch moves from the footer to the header, next to the theme toggle.
The site is served from computedtomography.org, now published from this repository.
Entry notes now translate. The 132 notes (and the Radon entry's) were rendered in English on the Chinese site even where a translation existed, because the site export read the English field directly and never consulted the overlay. Each language now builds its own payload, and build_site.py fails if a translated note does not reach that language's data.json.
— Static pages, a non-commercial content licence, and no more pull requests
Added (About, Contact and Privacy pages)
New static pages at /about/, /contact/ and /privacy/ (scope and verification method, how to suggest or correct an entry, what the site does and doesn't collect).
Compact footer on every page linking to them; licence line kept.
build_site.py renders the pages from web/pages/ and adds them to the sitemap.
Changed (maintenance)
The list is maintained by its authors only and no longer takes pull requests. CONTRIBUTING.md and the pull-request template are gone. The entry schemas and verification rules now live in MAINTENANCE.md#adding-entries. Suggestions come in as issues or by email.
Changed (licence)
Content is now CC BY-NC-SA 4.0 (was CC BY 4.0); code stays MIT. The maintainers want the list credited, kept non-commercial, and adaptations shared alike. LICENSE-content holds the full, unmodified legal code. Not retroactive: copies received earlier under CC BY 4.0 keep that licence.
Contact address added to the README and the site footer: heitieya@163.com.
First public state. The list is stable enough to cite.
Contents
289 papers spanning 1976–2026 across 17 topics, plus 43 toolkits, 34 datasets, 19 benchmarks/challenges and 29 learning resources.
Every paper link was resolved against Crossref, arXiv or the publisher's own record, and every URL was checked.
226 of 289 papers (78%) carry a peer-reviewed venue. Entries without one render as preprint year in italics, so the reviewed/unreviewed split is visible in the README itself rather than left for the reader to guess.
Added
The foundational papers of the three thinnest sections. Spectral, scatter and dynamic/4D held 6, 6 and 9 entries while other sections held 45–62, because they were missing the works the sub-fields are built on rather than because little exists. 17 papers recovered, taking them to 12, 11 and 15: Alvarez & Macovski 1976 (the founding spectral-CT paper), Roessl & Proksa 2007 and Schlomka 2008 (photon-counting K-edge imaging), Elbakri & Fessler 2002 and De Man 2001 (polyenergetic statistical reconstruction), Siewerdsen & Jaffray 2001 and Zhu 2006 (the CBCT scatter references), Sonke 2005 (the founding 4D-CBCT paper), Chen 2008 (PICCS), Leng 2008, Rit 2009, Bergner 2010 and Gao 2012.
The 2025–2026 frontier: 94 entries, weighted towards diffusion priors, Gaussian splatting, implicit neural representations and dynamic/4D CBCT.
MAINTENANCE.md, this changelog, an .editorconfig, and issue/PR templates.
The website source, which renders this same corpus as a searchable single page.web/ (template, stylesheet, client script), site.config.yml (title and the single url that every absolute address is derived from) and scripts/build_site.py produce a site/ directory; scripts/export_site_data.py produces the JSON payload on its own for anything that wants the data programmatically. The build reads data/*.yaml and calls build.py's own validator, so the page and the README cannot disagree, and it adds no dependencies. site/ is git-ignored and regenerated.
It is deployed: https://computedtomography.org/, published by .github/workflows/deploy.yml on every push to main. That workflow deliberately began as manual-only — actions/deploy-pages fails while Pages is switched off, and a workflow that goes red for reasons unrelated to the change under review teaches people to ignore red — and the push trigger was added once Pages was enabled. SITE.md documents the file map, the four Settings-side steps to reproduce the setup, the DNS records, the one-line domain switch, and the failure modes seen while writing it. CONTRIBUTING.md and MAINTENANCE.md pointed at it when it was written; CONTRIBUTING.md was dropped the same day and the README does not link it, so in practice MAINTENANCE.md is the surviving entry point.
The page itself: a hero, six stat tiles and a framed finder panel. The stat tiles read from the same counts as the README coverage line ({{N_PAPERS}}, {{PEER_PCT}}, …) rather than hardcoding numbers, so they cannot drift from data/. data-goto links on the tiles and the hero buttons switch to the matching tab; with JavaScript off they degrade to plain #tabs / #q anchors.
A second harvest pass: 32 papers recovered from a cross-check against two private papernotes collections. The two files held 43 works between them (39 ∪ 34, 30 in common); 10 were already on the list and 33 were new. 32 are added — the 33rd, a pure learning-theory paper that carries no paper link, was left out as out of scope. The additions are weighted to the newest venues (CVPR / ICCV / ECCV / ICLR / ICML / NeurIPS / AAAI / ICASSP 2024–2026, plus one journal) because that is where the collections were densest.
Four entries carry a paper link that is neither an arXiv id nor a DOI — the first in the list to do so. Three CVPR 2026 papers (CryoDeRec, Exact-GS, GH-NAF) have no arXiv posting and no registered DOI, so they link to their CVF OpenAccess page; MedGMAE (ICLR 2026) links to its OpenReview forum. Each was verified against the page's own citation_title metadata or the conference's own listing, and the coverage line in the README now says so rather than claiming every link came from Crossref or arXiv.
Fixed
scatter-rev pointed at the wrong half of a pair. The entry carried Part 1's title but Part 2's DOI (10.1118/1.3589140), so the link led to a different paper than the title promised. Both parts are now present and each points at its own DOI; the two subtitles ("Scatter compensation approaches" / "Scatter estimation approaches") are in the titles so the pair can be told apart.
Removed 88 duplicate entries. A harvesting pass had appended papers that were already on the list, under a different id: existing entries use short abbreviation ids (lpd, dip, super), harvested ones use title slugs, so the two never collided and duplicate detection — which compared id — saw nothing wrong. Detection now uses a canonical identity (arXiv id, DOI, or normalised title plus year), and build.py fails CI when the same work appears twice.
Normalised venue strings: 85 distinct spellings collapsed to 70. This repaired strings truncated at their source (the Journal of Machine Learning for Biomedical Imaging (MELB → MELBA), an unescaped HTML entity (Physics in Medicine & → PMB), synonym splits (IEEE Transactions on Medical Imaging vs TMI), years duplicated inside the venue (which the generated README printed twice), and six entries that carried a preprint server (arXiv, TechRxiv) where a venue belongs.
Corrected years that had recorded the arXiv posting year instead of the year at the venue, and one entry whose year and venue had been inherited from an unrelated paper that happened to share its short id.
Recovered 14 missing venues: 11 from the source records, and 3 more — Epi-NAF → ISBI 2025, Generative Modeling in Sinogram Domain → TRPMS 2024, CoCoDiff → SPIE 2022 — by checking candidate DOIs one at a time in Crossref. Two of those also exposed a wrong year that no title-based check would have caught.
Fixed 8 dead table-of-contents links and a broken badge.
Two wrong venues and one imprecise one, all inherited from the source notes.Regularizing INR with Diffusion Prior … Neutron CT was recorded as CVPR 2025 — it is ICASSP 2026 (10.1109/ICASSP55912.2026.11464479, pp. 21947–21951, Barcelona); Variational Garrote for Sparse Inverse Problems was recorded as CVPR 2025 — it is Journal of Computational Science 2026, vol. 100, art. 102980 (10.1016/j.jocs.2026.102980). Modeling X-ray Photon Pile-up with a Normalizing Flow was recorded as NeurIPS 2025, which is not wrong — it is the NeurIPS 2025 ML4PS workshop, so only the precision needed fixing.
The two CVPR claims were not simply overruled by a better source; the "a conference version came first, a journal version later" reading was ruled out explicitly, since that would have made the notes right rather than wrong. No CVPR record exists for either paper anywhere: not in the CVPR 2025 main proceedings (2,871 papers), not in any of that year's 52 CVF workshop volumes, and likewise nothing in CVPR 2026 (4,042 papers / 86 workshop volumes) or ICCV 2025 (2,701 / 76) — and OpenAlex, which indexes journals, conferences and workshops, lists exactly two works for the first paper (ICASSP 2026 + arXiv) and exactly three for the second (J. Comput. Sci. + SNU repository + arXiv), with no preprint–proceedings relation registered in Crossref. The "CVPR 2025" label traces to the note sites themselves, which carry it inconsistently next to a March-2026 arXiv id, and is mirrored verbatim across at least three of them.
No venue claim was carried over unchecked: each was re-derived from the arXiv comment field and re-confirmed against Crossref, OpenAlex or dblp. Every arXiv id was also read back through the arXiv API instead of being trusted as printed: 25 of 26 matched their entry's title at similarity 1.000, the last differing only in wording ("neutron computed tomography data" vs "neutron CT data").
MedGMAE was listed with no paper link. It is an accepted ICLR 2026 poster; the OpenReview record and the conference listing were both located.
One of the new links was dead on arrival, and the link sweep caught it.Efficient Unrolled Networks pointed at github.com/romainvo/efficientunrolling (404); the repository is romainvo/efficient-unrolling. This is the argument for running the sweep after any import rather than trusting the source's URL.
Changed
Content is now licensed CC BY 4.0; code stays MIT. The repository carried a single MIT licence over everything, which is the wrong licence for the content half: the awesome standard is explicit that code licences are not acceptable for list content ("We strongly recommend the CC0 license, but any Creative Commons license will work. Code licenses like MIT, BSD, Apache, GPL, etc, are not acceptable."). LICENSE-content now holds the full, unmodified CC BY 4.0 legal code and covers data/, README.md and the generated site; LICENSE continues to cover scripts/, web/ and .github/. The split is stated in the README, the site footer, the JSON-LD license field, a <link rel="license"> tag and CONTRIBUTING.md.
Chosen over CC0 because the list wants to be credited, and over CC BY-NC-SA — which a comparable project uses — because NC would exclude the industrial-CT and medical-device readers who most need this list, and ShareAlike would stop anyone merging these entries into a permissively licensed project. The change is not retroactive: what was already published under MIT stays MIT for whoever received it.
Clarified that the License column in the toolkit and dataset tables describes those projects, not this list. data/ was also checked for text derived from another list's notes — there is none; the only non-ASCII prose in it is nine Chinese author names (庄天戈, 张朝宗), and paper records carry no authors, no abstract and no field over 300 characters.
README.md is now generated from data/*.yaml. CI fails if the committed copy differs from a fresh build, so the README can no longer drift from the data.
Entries without a venue render as preprint 2025 in italics instead of a bare year, and the coverage line reports the reviewed/preprint split. (The wording was later shortened to the bare year inside the italics; see the 2026-10-11 entry.)
Renamed the repository to awesome-computed-tomography, and reworked the description, topics and the README scope paragraph so the project is findable by the vocabulary of the field rather than by the project name alone.
The GitHub topic set is now chosen from measured repository counts, not by keyword guesswork. With one star, only topics holding well under ~100 repositories can put the repository on page 1, so six niche tags the list documents were added (spectral-ct, dual-energy-ct, limited-angle, 4d-ct, scatter-correction, ct-imaging) and six unreachable or redundant ones dropped, notably medical-image-processing (931 repos, saying what medical-imaging already said) and awesome (11,730, redundant with awesome-list). Two spellings were rejected for matching zero repositories: limited-angle-tomography and dynamic-ct.