# Third-party notices

Dependency and source review: **15–17 October 2026**. The items below have separate terms. These notices do not declare a universal license for NeuroAtlas or its newly written application code; that license choice remains deferred. No endorsement by the named institutions or creators is implied.

## WebLLM runtime

- Project: [MLC-AI WebLLM](https://github.com/mlc-ai/web-llm).
- Selected version: **0.2.85**.
- Browser import: `https://cdn.jsdelivr.net/npm/@mlc-ai/web-llm@0.2.85/lib/index.js`.
- License: **Apache License 2.0**; [release license](https://raw.githubusercontent.com/mlc-ai/web-llm/v0.2.85/LICENSE), [local standard license text](licenses/Apache-2.0.txt).
- The release [prebuilt configuration](https://raw.githubusercontent.com/mlc-ai/web-llm/v0.2.85/src/config.ts) pairs model records with compiled libraries under `v0_2_84/base`. Preserve that pairing.

The release license additionally lists its bundled `src/openai_api_protocols` component under Apache License 2.0.

The runtime and associated artifacts are supplied by their original providers and fetched when the browser model is loaded. Preserve upstream copyright, attribution and applicable NOTICE material when redistributing or modifying a distribution. This register is not a substitute for retaining component-specific notices contained in the distributed runtime and its dependencies. No local modification to third-party runtime source is claimed here.

## Qwen model and MLC conversion

- Original: [Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct), by the **Qwen Team**.
- Original license: **Apache License 2.0**; [upstream LICENSE](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/blob/main/LICENSE), [local standard license text](licenses/Apache-2.0.txt).
- Original model copyright notice: **Copyright 2024 Alibaba Cloud**.
- Selected conversion: [mlc-ai/Qwen2.5-0.5B-Instruct-q4f16_1-MLC](https://huggingface.co/mlc-ai/Qwen2.5-0.5B-Instruct-q4f16_1-MLC), hosted by **MLC-AI**.
- Provenance: its [model card](https://huggingface.co/mlc-ai/Qwen2.5-0.5B-Instruct-q4f16_1-MLC/raw/main/README.md) identifies the original Qwen model and the `q4f16_1` MLC conversion format.
- Verified conversion revision: [`32ff081fe7e4dfe4ffb167b94c66fdf11e02b8ad`](https://huggingface.co/mlc-ai/Qwen2.5-0.5B-Instruct-q4f16_1-MLC/commit/32ff081fe7e4dfe4ffb167b94c66fdf11e02b8ad).

The converted repository does not display a separate model-license field in its current card; the license reference above is the original model's explicit Apache 2.0 license, with conversion provenance recorded separately. Do not describe the weights as authored by NeuroAtlas. If redistributing a local copy or further modification, retain the required original license and notices and identify modifications. The application does not include model-weight files in its static site package.

## Servier Medical Art illustration

Local asset: `assets/brain.png`, originally `cerveau_tronc.png`, from [Brain — Servier Medical Art](https://smart.servier.com/smart_image/smart-brain/) by **Servier Medical Art / Les Laboratoires Servier**. The image is licensed under **[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)**, independently of the runtime and model licenses.

Image provided by [Servier Medical Art](https://smart.servier.com/), licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

The source PNG was renamed without changing its content. NeuroAtlas adds approximate educational callout markers as a separate HTML overlay; they are not original Servier labels or anatomical boundary maps. See Servier's [attribution guidance](https://smart.servier.com/how-to-cite-servier-medical-art/) and [terms](https://smart.servier.com/terms-of-use/). Servier's website text, logos and trademarks are not covered by the medical-image license and were not copied into this application.

## Anatomy source information

The corpus reviewed on **16 October 2026** contains nineteen independently worded English passages for six regions, with seven institutional source records in `data/anatomy.json`. The twelve original location/function passage IDs and marker positions are preserved. Seven additions cover motor cortex, speech production, somatosensory cortex, hippocampal memory roles, visual cortex, coordination/posture and brainstem components:

- [Know Your Brain (Brain Basics)](https://www.ninds.nih.gov/health-information/public-education/brain-basics/brain-basics-know-your-brain), National Institute of Neurological Disorders and Stroke / NIH. [Reuse policy](https://www.ninds.nih.gov/publications/publications-help).
- [Get to Know Your Brain](https://www.nimh.nih.gov/news/media/2023/get-to-know-your-brain), National Institute of Mental Health / NIH. [Reuse policy](https://www.nimh.nih.gov/site-info/policies).
- [Definition of cerebellum](https://www.cancer.gov/publications/dictionaries/cancer-terms/def/cerebellum) and [Definition of brain stem](https://www.cancer.gov/publications/dictionaries/cancer-terms/def/brain-stem), National Cancer Institute / NIH. [Reuse policy](https://www.cancer.gov/policies/copyright-reuse).
- [Motor Cortex, MeSH D009044](https://meshb.nlm.nih.gov/record/ui?ui=D009044), [Visual Cortex, MeSH D014793](https://meshb.nlm.nih.gov/record/ui?ui=D014793), and [Hippocampus, MeSH D006624](https://meshb.nlm.nih.gov/record/ui?ui=D006624), MeSH 2026, National Library of Medicine / NIH. **Source: National Library of Medicine.** [NLM copyright policy](https://www.nlm.nih.gov/web_policies.html#copyright).

NINDS and NIMH policies identify their prepared information as public domain subject to stated exceptions. NCI permits reuse of its text unless otherwise indicated, with credit and links to the original titled products. NLM-produced U.S. government information is public domain unless otherwise indicated; its policy separately identifies third-party content exceptions. These text policies do not establish permission for every institutional image or logo and do not change the separate licenses of code or MRI assets. The descriptions are original short factual paraphrases, not copied MeSH database records, source code or imagery.

The NINDS [official PDF](https://www.ninds.nih.gov/sites/default/files/2025-05/know-your-brain-brian-basics.pdf), November 2023, NIH Publication No. 23-NS-3440a, was checked through official indexed text on printed pages 4 and 7. Live webpage/PDF access returned HTTP 403; no successful direct download is claimed. The NINDS publications-help policy was also checked through indexed text. Live NCI definitions/reuse policy and the three live NLM descriptor records were checked during the 16 October expansion. Hippocampal memory function is supported separately by NINDS.

No NIH images, source PDFs, course PDFs, full source datasets, or clinical CSV records are included. The limited MRI-derived PNG export has its own source and data license below. Source provenance and citation syntax do not establish that every generated statement follows from its sources.

The application guard added on 17 October 2026 checks sixteen known anatomical-name patterns against the text cited by each sentence and uses labeled source passages if validation fails. It introduces no third-party dependency or service and does not change the 16 October source review. This lexical check does not verify factual entailment. The separate project evaluation record describes the initial answer review and subsequent observed-failure regressions; neither is independent, human or clinical validation.

## Thesis architecture summary and training plots

The architecture descriptions and historical training plots relate to **Bogdan-Cristian Nuică's thesis project**, [source repository](https://gitlab.upt.ro/bogdan.nuica/licenta-nuica-bogdan-cristian-retele-neuronale-convolutionale-pe-imagini-rmn-asupra-creierului-uman). The static package includes six MRI-free loss/Dice plot PNGs extracted unchanged from notebook outputs: zero-based cell 16 in `unet.ipynb` and `segnet.ipynb`, and cell 13 in `Resnet50+Unet.ipynb`. Figure hashes and provenance are in `data/thesis.json`. Captions identify the original “Accuracy graph” as Dice coefficient. Loss uses blue validation/red training; Dice uses blue training/red validation.

The thesis repository declares Apache 2.0 for its code and credits **Copyright 2022 Emre Kele**, [upstream notebook](https://www.kaggle.com/code/emrekele/braintumordetection). Its data-generator provenance also cites the MIT-licensed [zhixuhao/unet](https://github.com/zhixuhao/unet) project. No training-generator source was copied into the current website or newly written local loader. These attributions do not assign rights to third-party MRI data or recovered model weights.

The notebooks reference [Kaggle: LGG MRI Segmentation](https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation), associated with [TCIA: TCGA-LGG](https://www.cancerimagingarchive.net/collection/tcga-lgg/). The package now includes the limited MRI/mask adaptation described below; recovered weights and full source data are not included. Present TCIA access restrictions alone do not determine the permissions of older derivatives. Training plots describe historical runs; saved best-checkpoint test metrics are not reported as measurements of the recovered final-weight files.

**Dataset review, 16 October 2026:** Mateusz Buda's [Kaggle metadata](https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation/metadata) and official dataset API identify version 2 as **[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/)**. It supports the documented noncommercial educational recovery checks and the limited saved-output illustrations below. Credit the creator/source, retain the license, identify changes, and meet ShareAlike conditions when sharing adaptations. These data rights are distinct from the thesis code license. Original training overlap is unknown; no independent accuracy claim follows from reuse. Trained-weight distribution remains separately reviewed.

The private cohort comparison produced local HTML/CSV reports and composite MRI figures, which remain outside this package. A separate exporter supplies only the six-example assets listed below. Full-cohort records, original TIFFs, prediction caches, and recovered model weights are not bundled.

## LGG MRI examples and saved prediction masks

- Source author: **Mateusz Buda**.
- Source release: **LGG Segmentation Dataset / Brain MRI segmentation**, [Kaggle version 2](https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation), derived from TCIA imaging.
- License for the data-derived PNGs: **[Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International](https://creativecommons.org/licenses/by-nc-sa/4.0/)**.
- Included scope: thirty 256 × 256 grayscale PNGs under `assets/comparison/`: six FLAIR images, six reference masks, and eighteen saved prediction masks; **153,232 bytes total** at this checkpoint. `data/comparison.json` supplies example aliases, selected-example scores, cohort summaries, source details, and asset hashes.
- Citation: Buda M., Saha A., Mazurowski M. A. *Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm*. Computers in Biology and Medicine 109 (2019). [DOI: 10.1016/j.compbiomed.2019.05.002](https://doi.org/10.1016/j.compbiomed.2019.05.002).
- Changes: FLAIR-channel extraction and PNG conversion; reference and fixed-threshold (`> 0.5`) prediction masks exported as binary PNGs; NeuroAtlas adds colored display overlays. There is no spatial cropping or rescaling. Predictions were produced by the recovered thesis models.
- Retained notice: [MRI comparison asset notice](assets/comparison/NOTICE.txt). Attribution and change information are also embedded in the fresh PNG metadata and shown in the interface. Raw patient identifiers, dates, and private paths are omitted.

Use the licensed material noncommercially, credit the creator/source and license, identify modifications, and retain the required license when sharing adaptations. These conditions do not automatically license separate application code or recovered model weights. No endorsement by the dataset authors or TCIA is implied.

The six examples were fixed before scoring and come from three patient folders in the original cohort with unknown training overlap. They are educational illustrations, not independent-test or clinical validation. The **Compare MRI** section reads precomputed outputs without new inference, image selection, or segmentation-weight downloads. All six exported FLAIR images are nonblank; the viewer also preserves and explains a blank source channel. That saved-comparison section passed 62/62 installed Chrome 155 checks covering image/score consistency, overlay controls, scoped zoom, responsive containment, race handling, and failure retries without weight/private-data downloads. Older thesis/private-report results describe separate checks. Other browsers and mobile GPUs were not tested by that run. The separate **Try a slice** feature below performs browser-local inference. The app has not been deployed in this checkpoint.

Local model checks use Keras 3.15.1 (Apache 2.0) and h5py 3.16.0 (BSD-3-Clause), with license files retained in their installed distributions. These are local Python tools, not libraries bundled in this static website. All three models passed synthetic execution checks, and the SegNet local prediction/export path passed on a synthetic RGB fixture. The static thesis section passed 38 Chrome checks, including all six plots and resource-failure recovery, without AI requests or model-weight downloads. These checks do not establish TensorFlow/Torch parity, real-MRI quality, or clinical performance.

Separate private conversion checks use ONNX 1.20.1 (Apache-2.0), ONNX Runtime 1.31.0 (MIT), FlatBuffers 25.12.19 (Apache 2.0), and protobuf 5.29.6 (BSD-3-Clause), as identified in installed package metadata. The isolated QA files and ONNX model files remain outside the static website package. All three private exports passed 15 native and 30 browser synthetic inference cases, with consistency against recovered Keras/Torch outputs and zero thresholded-mask differences. Original TensorFlow parity and independent real-MRI quality remain unverified; conversion does not establish model redistribution rights.

## ONNX Runtime Web and prepared-slice inference

- Provider: **Microsoft Corporation / ONNX Runtime**.
- Selected browser version: **1.30.0**.
- Source: [versioned ONNX Runtime Web project](https://github.com/microsoft/onnxruntime/tree/v1.30.0/js/web).
- Browser bundle: `https://cdn.jsdelivr.net/npm/onnxruntime-web@1.30.0/dist/ort.webgpu.min.js`, with matching WASM assets under that same version's `dist/` prefix.
- License: **MIT**, **Copyright (c) Microsoft Corporation**. Retained [local license](licenses/ONNXRuntime-MIT.txt); [versioned upstream license](https://raw.githubusercontent.com/microsoft/onnxruntime/v1.30.0/LICENSE).
- Additional upstream notices: [versioned ThirdPartyNotices.txt](https://raw.githubusercontent.com/microsoft/onnxruntime/v1.30.0/ThirdPartyNotices.txt). Preserve applicable notices when copying, modifying, or redistributing runtime components.
- Review date: **17 October 2026**. No modification to third-party runtime source is claimed.

The runtime is used by the isolated QA and the main-app **Try a slice** feature. It loads from its CDN only after explicit model loading. Model weights remain outside `dist`: the local server can expose exact verified model GET routes only when started with `--enable-local-mri`, or a user may select a matching local ONNX file in the browser. The default server exposes only `dist`. These local workflows do not settle permission for public weight distribution or hosting. The runtime's MIT license does not license the weights, dataset, or user images.

The new feature accepts prepared native 256 × 256 MRI-modality PNGs, uses a strict `probability > 0.5` mask threshold, and runs one selected model in browser memory. Missing optional modality channels are explicitly substituted with FLAIR. It performs no resizing, cropping, registration, or intensity normalization beyond `/255`. Automatic compute tries WebGPU and permits CPU/WASM fallback; some WebGPU-session operations may run on the CPU. User images, optional selected ONNX files, and predictions are not POSTed or sent to an inference API. CDN and hosting requests still disclose ordinary request metadata. No paid service or public deployment is introduced.

The preparation and input checks establish accepted-pixel handling, not diagnosis or independent accuracy. A model mask is an educational prediction; an empty mask does not rule out an abnormality. New-patient performance remains unknown, and no Dice/IoU is inferred for user images without reference annotations.

## Structured source selection

Reviewed on **20 October 2026**. Source-answer selection uses the existing WebLLM **0.2.85** runtime's [JSON-schema response format](https://github.com/mlc-ai/web-llm/blob/v0.2.85/src/openai_api_protocols/chat_completion.ts). Its [release package](https://github.com/mlc-ai/web-llm/blob/v0.2.85/package.json) pins **@mlc-ai/web-xgrammar 0.1.27**. XGrammar's [versioned license](https://github.com/mlc-ai/xgrammar/blob/v0.1.27/LICENSE) is **Apache 2.0**, retained through the existing [local Apache license](licenses/Apache-2.0.txt). Credit the MLC-AI/XGrammar contributors. No third-party source modification, new hosted inference service, or additional model weights are introduced by this response-format change. Schema compliance controls the shape of a model response; it does not establish factual relevance or completeness.

## English question-framing lexicon

The RAG question-framing helper includes the 318 English stop words exported from the existing **scikit-learn 1.9.1** installation (sklearn.feature_extraction.text.ENGLISH_STOP_WORDS). Copyright (c) 2007–2026 The scikit-learn developers; **BSD 3-Clause**. The applicable copyright, conditions, and disclaimer are retained in [ScikitLearn-BSD-3-Clause.txt](licenses/ScikitLearn-BSD-3-Clause.txt). No scikit-learn binaries or bundled third-party runtimes are distributed here.

Only an explicitly reviewed grammatical subset is eligible for question framing. The application retains known source content and meaning-changing qualifiers when applying it. The helper performs limited vocabulary-aware English inflection matching; it is not a full lemmatizer, embedding model, or factual-entailment check. It does not alter the raw TF-IDF index or its Python/browser preprocessing.

## Optional MiniLM semantic encoder

Reviewed on **18 October 2026**. The browser semantic-search worker uses [Xenova/all-MiniLM-L6-v2](https://huggingface.co/Xenova/all-MiniLM-L6-v2/blob/751bff37182d3f1213fa05d7196b954e230abad9/README.md), the ONNX conversion of [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2). The converted and original model cards declare **Apache License 2.0**; retain the [local Apache license](licenses/Apache-2.0.txt). Credit the original sentence-transformers contributors and Xenova's ONNX conversion. No model authorship or endorsement by NeuroAtlas is implied.

The encoder revision is fixed to **751bff37182d3f1213fa05d7196b954e230abad9**. Its q8 `onnx/model_quantized.onnx` file is **22,972,370 bytes**, SHA-256 **afdb6f1a0e45b715d0bb9b11772f032c399babd23bfc31fed1c170afc848bdb1**. The tokenizer JSON adds **711,661 bytes**. Configuration and tokenizer files are also fixed to this revision. NeuroAtlas verifies their reviewed sizes and SHA-256 values, including cached copies, before loading. No encoder weights are bundled in the static package, and no modification to the converted weights or tokenizer is claimed.

The optional runtime is **Hugging Face Transformers.js 3.8.1**, fetched from `https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.8.1/dist/transformers.min.js` with matching versioned WASM files. Its Apache 2.0 release license is retained in [TransformersJS-3.8.1-LICENSE.txt](licenses/TransformersJS-3.8.1-LICENSE.txt). The release pins **ONNX Runtime Web 1.22.0-dev.20250409-89f8206ba4**; its Microsoft MIT license and complete upstream dependency notices are retained in [ONNXRuntime-TransformersJS-MIT.txt](licenses/ONNXRuntime-TransformersJS-MIT.txt) and [ONNXRuntime-TransformersJS-ThirdPartyNotices.txt](licenses/ONNXRuntime-TransformersJS-ThirdPartyNotices.txt), from commit **89f8206ba4f1c22c39e0297fb55272e8ce8cd7d0**. This dependency is isolated from the separate MRI worker's ONNX Runtime 1.30.0. No third-party runtime source modification is claimed.

The facade creates its worker and downloads only after explicit loading. It uses one CPU/WASM thread, 256-token truncation, attention-mask mean pooling, L2-normalized 384-dimensional vectors, and cosine ranking of the local anatomy corpus. It uses no API key, account, paid inference service, or server-side question processing. Fixed CDN/model requests disclose ordinary network metadata; browser caching is best effort. Questions and vectors are processed in worker memory and are not stored by this feature. The worker closes its fetch path after indexing. Similarity scores do not establish that a question is answerable or that a generated answer is correct.

This is an educational demonstration, not a medically validated or personalized medical service. Source attribution and citation formatting do not guarantee that generated statements are correct or fully supported.