A human-in-the-loop classifier reduces non-tumor-cell gallery review
Source check: Source checked. Not independent scientific confirmation.
Finding: A self-supervised image encoder and clinician-updated SVM were evaluated on additional blood-image data from 10 metastatic breast-cancer patients. Their proposed gallery contained higher fractions of expert-identified circulating tumor cells than CellSearch galleries, while each system found cells the other missed. This does not establish complete sensitivity or clinical benefit.
Technology: DINO self-supervised XCiT encoder, SVM and cluster-targeted human-in-the-loop labeling
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- The trained encoder represented cell images and the SVM prioritized candidates after targeted human relabeling; the qualifying result is the additional-patient gallery/reference comparison, not a generic sampling benchmark.
- Test and result
- The 90-patient archive supplied 60 patients for self-supervised training, 20 for downstream development/testing, and 10 additional patients for final comparison. In those 10, candidate-gallery positive predictive values ranged from .479 to .878 for the final HiL pipeline versus .094 to .365 for CellSearch events. Actual CTC counts were similar overall, but CellSearch-only cells averaged 7 per patient and HiL-only cells averaged 6.
- Limitation / still unknown
- One expert reviewed both galleries, with blinding not stated; these labels are not an independent molecular ground truth. Selection of the uncertain development cluster used test-set performance before new examples were labeled, so development-test improvement is not untouched confirmation. The separate final 10-person comparison is small and selected; each pipeline missed cells found by the other, and complete CTC sensitivity is unknown. Higher gallery PPV does not establish improved diagnosis, survival or treatment selection. This extends prior learned cell-image representation work, rather than independently discovering a new encoder.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC12191740/
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: Original computational encoder/classifier Methods, patient-level partitions, human-labeling experiments and final 10-patient Results/Table2 actually read..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 1/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.