30-Second Takeaway
- MLH1 immunostaining can uncover morphologically subtle dysplasia in sessile serrated lesions when used selectively.
- Claudin-18 loss is a practical ancillary marker for non-mucinous lung adenocarcinoma in small or challenging biopsies.
- Multimodal and foundation-model AI approaches are nearing clinically useful performance for cancer subtyping and molecular surrogates.
- Pediatric cancers are strongly driven by structural variants, underscoring the need for improved SV detection workflows.
- Long-read sequencing, T2T references, and methylation or SV atlases are reshaping what is detectable in molecular diagnostics.
Week ending March 14, 2026
Emerging tools refining diagnostic precision in pathology, from serrated lesions and lung adenocarcinoma to gliomas and structural variants
MLH1 immunostaining modestly increases detection of dysplasia in sessile serrated lesions
In this nationwide Dutch series, dysplasia was identified in 9.4% of 186,427 sessile serrated lesions (SSLs). A national audit found diagnostic discrepancies in 11% of SSLs with dysplasia, while only 20% of laboratories used immunohistochemistry. MLH1 was the only biomarker with clear added value, increasing SSLd diagnoses from 33 to 41 and 35 to 43 in two cohorts. BRAF immunostaining of 1,572 advanced adenomas showed misdiagnosed SSLd among conventional adenomas was very rare. The authors recommend MLH1 IHC for selected morphologically suspicious SSLs to unmask dysplasia that does not meet overt criteria.
Claudin-18 loss is a highly accurate marker for non-mucinous lung adenocarcinoma
This study evaluated Claudin-18 (Cldn18) immunostaining in 391 lung resections and 53 small biopsies across lung cancer subtypes. Normal alveoli and reactive pneumocytes consistently expressed Cldn18, whereas non-mucinous lung adenocarcinomas (NM-LUADs) typically showed marked loss. Complete absence of Cldn18 occurred in 83% of NM-LUAD resections and 84.8% of NM-LUAD biopsies, with AUC 0.992, sensitivity 98.4%, specificity 100%. Lepidic NM-LUAD and atypical adenomatous hyperplasia often showed low, rather than preserved, expression, while many mucinous LUADs retained staining. The authors propose Cldn18 loss as an ancillary tool to separate NM-LUAD from mucinous tumors and reactive mimics, particularly in small specimens.
Multimodal large language models provide traceable pathology reasoning with high accuracy
This work reconfigures off-the-shelf multimodal large language models to perform stepwise diagnostic reasoning on breast and prostate pathology slides. Using small labeled subsets, a two-phase self-learning process derived diagnostic criteria without updating model weights. The framework exceeded 90% accuracy distinguishing normal tissue from invasive carcinoma while generating audit-ready, feature-linked rationales. It also differentiated entities such as ductal carcinoma in situ by autonomously highlighting nuclear atypia and architectural disruption. Expert pathologists judged the AI-generated descriptions as broadly concordant with accepted diagnostic criteria, supporting potential use for transparent decision support.
References
Numbered in order of appearance. Click any reference to view details.
Additional Reads
Optional additional studies from this edition.