{"id":"b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a","title":"Hernandez and coauthors: Where slide-level classifiers fail on atypical breast lesions","argument":"Nerea Hernandez and coauthors report that slide-level classifiers built on pathology foundation models perform worst at distinguishing atypical breast lesions, a distinction they say increasingly determines surgery. Their analysis points to a mismatch between slide labels and the lesion being assessed: slides were labelled by their most severe lesion, but only 39% of slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little. By contrast, classifiers trained per decision with lesion-level labels improved all four atypical boundaries. The authors also report that inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions, and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22. Escape Meridian's interpretation is that these findings suggest how training labels are assigned may matter alongside directing a classifier’s attention to a lesion. The abstract reports these results but does not establish how the approach would perform in wider clinical use.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"Evidence is the arXiv abstract, not the full paper; details needed to assess the study design and reported results are not available here. The abstract reports results on BRACS, a hospital setting and HISTAI, but does not establish performance in wider clinical use.","speaker":{"name":"Hernandez, Nerea, Herrera, Luis J., Ramirez, Cesar, Valenzuela, Olga, Rojas, Ignacio, Ortuño, Francisco M.","roleAtSource":"Article author"},"source":{"title":"Where slide-level foundation-model classifiers fail on atypical breast lesions","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.19630","publishedAt":"2026-11-14","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors report that slide-level classifiers built on pathology foundation models perform worst on atypical breast lesions, whose classification they say increasingly determines surgery.","excerpt":"Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Because slides were labelled by their most severe lesion, only 39% of slides with annotated ADH were labelled ADH.","excerpt":"Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Attention ranked the lesion high but rarely first; showing the model only the lesion helped little, while classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","excerpt":"Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia.","excerpt":"ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"The authors report lower triage specificity in a hospital dataset and HISTAI, where benign slides were scored closer to lesions.","excerpt":"In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"The reported findings raise the possibility that label assignment, not just which image region a classifier attends to, affects its atypical-lesion classifications.","excerpt":"showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","url":"https://arxiv.org/abs/2611.19630","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.19630","checkedAt":"2026-11-17T13:49:00.244Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Would classifiers trained per decision with lesion-level labels retain their improvement across clinical settings beyond those reported in the abstract?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Hernandez and coauthors: Where slide-level classifiers fail on atypical breast lesions","summary":"Nerea Hernandez and coauthors report that slide-level classifiers built on pathology foundation models perform worst at distinguishing atypical breast lesions, a distinction they say increasingly determines surgery. Their analysis points to a mismatch between slide labels and the lesion being assessed: slides were labelled by their most severe lesion, but only 39% of slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little. By contrast, classifiers trained per decision with lesion-level labels improved all four atypical boundaries. The authors also report that inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions, and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22. Escape Meridian's interpretation is that these findings suggest how training labels are assigned may matter alongside directing a classifier’s attention to a lesion. The abstract reports these results but does not establish how the approach would perform in wider clinical use.","question":"Would classifiers trained per decision with lesion-level labels retain their improvement across clinical settings beyond those reported in the abstract?","claims":[{"text":"The authors report that slide-level classifiers built on pathology foundation models perform worst on atypical breast lesions, whose classification they say increasingly determines surgery.","excerpt":"Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions.","support":"source-stated"},{"text":"Because slides were labelled by their most severe lesion, only 39% of slides with annotated ADH were labelled ADH.","excerpt":"Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH.","support":"source-stated"},{"text":"Attention ranked the lesion high but rarely first; showing the model only the lesion helped little, while classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","excerpt":"Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","support":"source-stated"},{"text":"Inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia.","excerpt":"ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia.","support":"source-stated"},{"text":"The authors report lower triage specificity in a hospital dataset and HISTAI, where benign slides were scored closer to lesions.","excerpt":"In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.","support":"source-stated"},{"text":"The reported findings raise the possibility that label assignment, not just which image region a classifier attends to, affects its atypical-lesion classifications.","excerpt":"showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","support":"interpretation"}],"limitations":["Evidence is the arXiv abstract, not the full paper; details needed to assess the study design and reported results are not available here.","The abstract reports results on BRACS, a hospital setting and HISTAI, but does not establish performance in wider clinical use."],"mediaId":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","correctionReason":null,"id":"b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a","publicationId":"6df47b2c-da43-4982-87af-00ea1bc9b92c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.19630","sourceTitle":"Where slide-level foundation-model classifiers fail on atypical breast lesions","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-14","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:28:43.804Z","createdAt":"2026-11-17T13:32:35.017Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","kind":"article","url":"https://arxiv.org/abs/2611.19630","sourceUrl":"https://arxiv.org/abs/2611.19630","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T13:31:30.725Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.19630","checkedAt":"2026-11-17T13:49:00.244Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","confidence":0.67,"provider":"typesafe","model":"jev-1.13.0","inputHash":"28127efa51537598d5bd7c37215984394f147d465ebba4c485b7ccd6a980872a","decidedAt":"2026-11-17T13:31:32.021Z"},"provenance":{"textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observationId":"499e1906251a4fe06285ced422d2b12d0b5b4305fba9a5676c31cc0b8be2cf2b","retrievedAt":"2026-11-17T13:31:30.561Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"4f495dbf-d82a-440a-8469-8d9ab9bdecf9"},"reviewer":{"model":"gpt-6-luna","runId":"c695f4d5-96e9-42d6-8fba-cbe14436db28"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T13:32:35.017Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hernandez, Nerea","url":null,"kind":"unknown"},{"name":"Herrera, Luis J.","url":null,"kind":"unknown"},{"name":"Ramirez, Cesar","url":null,"kind":"unknown"},{"name":"Valenzuela, Olga","url":null,"kind":"unknown"},{"name":"Rojas, Ignacio","url":null,"kind":"unknown"},{"name":"Ortuño, Francisco M.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hernandez, Nerea","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hernandez, Nerea","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Herrera, Luis J.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Herrera, Luis J.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Ramirez, Cesar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ramirez, Cesar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Valenzuela, Olga","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Valenzuela, Olga","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Rojas, Ignacio","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Rojas, Ignacio","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Ortuño, Francisco M.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ortuño, Francisco M.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.19630","snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a/source"}},"canonicalUrl":"https://escapemeridian.com/readings/hernandez-and-coauthors-where-slide-level-classifiers-fail-on-b1152b57","formats":{"html":"/readings/hernandez-and-coauthors-where-slide-level-classifiers-fail-on-b1152b57","markdown":"/readings/hernandez-and-coauthors-where-slide-level-classifiers-fail-on-b1152b57.md","json":"/readings/hernandez-and-coauthors-where-slide-level-classifiers-fail-on-b1152b57.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-foundation-model","name":"Foundation model","explanation":"A large model trained on broad data that can be adapted to many different tasks.","referenceUrl":"https://developers.google.com/machine-learning/glossary#foundation-model"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a","revision":"6df47b2c-da43-4982-87af-00ea1bc9b92c","revisionKind":"publication-id","title":"Hernandez and coauthors: Where slide-level classifiers fail on atypical breast lesions","url":"https://escapemeridian.com/readings/hernandez-and-coauthors-where-slide-level-classifiers-fail-on-b1152b57"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.19630","revision":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","revisionKind":"sha256-source-text","title":"Where slide-level foundation-model classifiers fail on atypical breast lesions","url":"https://arxiv.org/abs/2611.19630"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a","title":"Hernandez and coauthors: Where slide-level classifiers fail on atypical breast lesions","argument":"Nerea Hernandez and coauthors report that slide-level classifiers built on pathology foundation models perform worst at distinguishing atypical breast lesions, a distinction they say increasingly determines surgery. Their analysis points to a mismatch between slide labels and the lesion being assessed: slides were labelled by their most severe lesion, but only 39% of slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little. By contrast, classifiers trained per decision with lesion-level labels improved all four atypical boundaries. The authors also report that inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions, and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22. Escape Meridian's interpretation is that these findings suggest how training labels are assigned may matter alongside directing a classifier’s attention to a lesion. The abstract reports these results but does not establish how the approach would perform in wider clinical use.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"Evidence is the arXiv abstract, not the full paper; details needed to assess the study design and reported results are not available here. The abstract reports results on BRACS, a hospital setting and HISTAI, but does not establish performance in wider clinical use.","speaker":{"name":"Hernandez, Nerea, Herrera, Luis J., Ramirez, Cesar, Valenzuela, Olga, Rojas, Ignacio, Ortuño, Francisco M.","roleAtSource":"Article author"},"source":{"title":"Where slide-level foundation-model classifiers fail on atypical breast lesions","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.19630","publishedAt":"2026-11-14","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors report that slide-level classifiers built on pathology foundation models perform worst on atypical breast lesions, whose classification they say increasingly determines surgery.","excerpt":"Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Because slides were labelled by their most severe lesion, only 39% of slides with annotated ADH were labelled ADH.","excerpt":"Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Attention ranked the lesion high but rarely first; showing the model only the lesion helped little, while classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","excerpt":"Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"Inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia.","excerpt":"ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"The authors report lower triage specificity in a hospital dataset and HISTAI, where benign slides were scored closer to lesions.","excerpt":"In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.","url":"https://arxiv.org/abs/2611.19630","relationship":"Source-stated"},{"text":"The reported findings raise the possibility that label assignment, not just which image region a classifier attends to, affects its atypical-lesion classifications.","excerpt":"showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","url":"https://arxiv.org/abs/2611.19630","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.19630","checkedAt":"2026-11-17T13:49:00.244Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Would classifiers trained per decision with lesion-level labels retain their improvement across clinical settings beyond those reported in the abstract?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Hernandez and coauthors: Where slide-level classifiers fail on atypical breast lesions","summary":"Nerea Hernandez and coauthors report that slide-level classifiers built on pathology foundation models perform worst at distinguishing atypical breast lesions, a distinction they say increasingly determines surgery. Their analysis points to a mismatch between slide labels and the lesion being assessed: slides were labelled by their most severe lesion, but only 39% of slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little. By contrast, classifiers trained per decision with lesion-level labels improved all four atypical boundaries. The authors also report that inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions, and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22. Escape Meridian's interpretation is that these findings suggest how training labels are assigned may matter alongside directing a classifier’s attention to a lesion. The abstract reports these results but does not establish how the approach would perform in wider clinical use.","question":"Would classifiers trained per decision with lesion-level labels retain their improvement across clinical settings beyond those reported in the abstract?","claims":[{"text":"The authors report that slide-level classifiers built on pathology foundation models perform worst on atypical breast lesions, whose classification they say increasingly determines surgery.","excerpt":"Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions.","support":"source-stated"},{"text":"Because slides were labelled by their most severe lesion, only 39% of slides with annotated ADH were labelled ADH.","excerpt":"Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH.","support":"source-stated"},{"text":"Attention ranked the lesion high but rarely first; showing the model only the lesion helped little, while classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","excerpt":"Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","support":"source-stated"},{"text":"Inserting ADH regions into benign slides raised the probability of malignancy more than the probability of atypia.","excerpt":"ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia.","support":"source-stated"},{"text":"The authors report lower triage specificity in a hospital dataset and HISTAI, where benign slides were scored closer to lesions.","excerpt":"In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.","support":"source-stated"},{"text":"The reported findings raise the possibility that label assignment, not just which image region a classifier attends to, affects its atypical-lesion classifications.","excerpt":"showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries.","support":"interpretation"}],"limitations":["Evidence is the arXiv abstract, not the full paper; details needed to assess the study design and reported results are not available here.","The abstract reports results on BRACS, a hospital setting and HISTAI, but does not establish performance in wider clinical use."],"mediaId":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","correctionReason":null,"id":"b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a","publicationId":"6df47b2c-da43-4982-87af-00ea1bc9b92c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.19630","sourceTitle":"Where slide-level foundation-model classifiers fail on atypical breast lesions","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-14","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:28:43.804Z","createdAt":"2026-11-17T13:32:35.017Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","kind":"article","url":"https://arxiv.org/abs/2611.19630","sourceUrl":"https://arxiv.org/abs/2611.19630","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T13:31:30.725Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.19630","checkedAt":"2026-11-17T13:49:00.244Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"11350f8b502438353820a9363200d5701bf9e8c5e7be25637baa0e878f13e9e7","confidence":0.67,"provider":"typesafe","model":"jev-1.13.0","inputHash":"28127efa51537598d5bd7c37215984394f147d465ebba4c485b7ccd6a980872a","decidedAt":"2026-11-17T13:31:32.021Z"},"provenance":{"textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observationId":"499e1906251a4fe06285ced422d2b12d0b5b4305fba9a5676c31cc0b8be2cf2b","retrievedAt":"2026-11-17T13:31:30.561Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"4f495dbf-d82a-440a-8469-8d9ab9bdecf9"},"reviewer":{"model":"gpt-6-luna","runId":"c695f4d5-96e9-42d6-8fba-cbe14436db28"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T13:32:35.017Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hernandez, Nerea","url":null,"kind":"unknown"},{"name":"Herrera, Luis J.","url":null,"kind":"unknown"},{"name":"Ramirez, Cesar","url":null,"kind":"unknown"},{"name":"Valenzuela, Olga","url":null,"kind":"unknown"},{"name":"Rojas, Ignacio","url":null,"kind":"unknown"},{"name":"Ortuño, Francisco M.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hernandez, Nerea","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hernandez, Nerea","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Herrera, Luis J.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Herrera, Luis J.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Ramirez, Cesar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ramirez, Cesar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Valenzuela, Olga","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Valenzuela, Olga","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Rojas, Ignacio","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Rojas, Ignacio","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}},{"actor":{"name":"Ortuño, Francisco M.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ortuño, Francisco M.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.19630","locator":"head/meta[name=author|citation_author]"},"assertionId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T13:49:00.244Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","sourceUrl":"https://arxiv.org/abs/2611.19630","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.19630","snapshotId":"018d9a7f0b19070a41965cf6c8f965c07630246deaba6c08b12aec009f35d372","contentSha256":"1bcfcd2fe1e6da5e40d54e6880cd841527f00c87b681377d9af4c5a96c1912e9","textHash":"dcd2ca19cb394f0bee19a2371ae770a9c607208abf9ed2a01929dee3b0e77614","observedAt":"2026-11-17T13:49:00.244Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/b1152b57cdc0219f66e75f6d4c0499b2b657e7a957ab9c2a69531a3eaed02f8a/source"}}}],"audits":[{"publicationId":"6df47b2c-da43-4982-87af-00ea1bc9b92c","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft attributes the findings to Nerea Hernandez and coauthors, who are listed as authors, and keeps the results framed as their report. The Escape Meridian inference is clearly labeled as interpretation.","excerpt":"Authors:Nerea Hernandez, Luis J. Herrera, Cesar Ramirez, Olga Valenzuela, Ignacio Rojas, Francisco M. Ortuño"},{"dimension":"summary","status":"uncertain","explanation":"The draft is faithful but mostly repeats the abstract's technical wording. A general reader is not told what pathology foundation models, model attention, the four atypical boundaries, BRACS, HISTAI or triage specificity mean. It also does not explain why a fall to 0.00-0.07 specificity matters (many benign slides flagged). Smallest repair: add tooltips or brief plain-language glosses for these terms. Keep the technical terms exact.","excerpt":"triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22"},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-11-14 matches the v1 submission date. It is not the later feed date of 2026-11-17.","excerpt":"[Submitted on 14 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each factual claim closely tracks the abstract. The interpretation that label assignment may matter alongside attention is labeled, conditional and supported by the contrast the abstract draws. The limitations and the question do not overstate clinical benefit.","excerpt":"showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-17T13:32:35.017Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}