Paper · S7570
MAPS: pathologist-level cell type annotation from tissue images through machine learning
MAPS uses expert-labelled protein-expression data to classify cells. The authors report strong results within datasets, while acknowledging leakage risk from splitting one patient’s CODEX image by cells. Performance falls across cohorts and competing methods sometimes do better. Nolan is a coauthor and supervisor; disclosed commercial interests, reporting differences and missing original materials accompany the findings.
Read the paper at ebi.ac.ukWhat we checked
All 60,563 decoded characters of the supplied article XML were read in eight complete bounded returns, including 32 references. Root independently matched every returned character and verified 15 native authors. Four main JPEG figures and 17 equation graphics were viewed. All 50 supplied supplement pages were read in text and pixels: 20 supplementary-information pages with four figures and fifteen tables, 26 peer-review pages and four reporting pages. A truncated image return was excluded and replaced by small complete returns. Smallest Figure 1C scalar labels remain unresolved. Original main PDF, raw source data, code execution and the earlier preprint were not received or reviewed. The preprint belongs to the same work family.
Complete supplied article and supplement text read; original and tiny-label gaps
How we check sources
Not checked yet
- Expert-labelled cells, folds and reused datasets are dependent observations. Single-patient CODEX cell splitting can inflate performance. Transfer results and class-level comparisons do not establish uniform superiority, general clinical performance or independent biological truth.
- The model excludes spatial features and retains manual annotation and correction. Faster runtime does not mean every hardware measure improves; plotted RAM use exceeds a comparator. Critical and favorable reviews are preserved; acceptance is not independent reproduction.
- Results report cHL2 192795 cells versus over230895 in Methods. AP ranges differ from supplementary micro-AP values. Reporting Summary names MARIO, states no exclusions despite marker selection and labels a CODEX-style panel MIBI. A catalogue field is blank; causes and source-version equivalence remain unresolved.
- Nolan’s declared role is supervision and manuscript review, without specific model-development or pathological-annotation credit. IonPath and Akoya interests, patent and advisory membership are disclosed. No UAP cohort, government approach, secret programme or clinical validation occurs in this paper.
- Original main PDF, tiny Figure 1C scalars, raw channels, masks, Source Data, code and preprint history remain unresolved or unreviewed. Supplement matrices do not automatically repair exact ambiguous main-panel values. Public aggregate publication reading is distinct from private patient-record collection.
Cite this record
Muhammad Shaban, Yunhao Bai, Huaying Qiu, Shulin Mao, Jason Yeung, Yao Yu Yeo, Vignesh Shanmugam, Han Chen, Bokai Zhu, Jason L Weirather, Garry P Nolan, Margaret A Shipp, Scott J Rodig, Sizun Jiang, Faisal Mahmood. MAPS: pathologist-level cell type annotation from tissue images through machine learning. January 2, 2024. BioParity S7570. https://bioparity.io/sources/S7570Record data (JSON)