Shi, Qiuying

Abstract Large language models (LLMs) and vision‐language models represent a fundamentally different category of artificial intelligence (AI) compared to prior image analysis approaches in digital pathology, which have largely been based on convolutional neural network architectures. This review from the American Society of Cytopathology Clinical Practice Committee examines the current evidence for LLM and vision‐language model applications in cytopathology, including structured reporting, diagnostic assistance, quality control, education, and workflow integration. The distinction between applications with preliminary evidence and those that remain hypothetical is described. A detailed assessment of the challenges that must be addressed before clinical deployment, including hallucination risk, limited explainability, bias, data privacy, validation gaps, and infrastructure barriers is discussed. A review of the regulatory landscape in the United States and European Union as it applies to AI‐enabled software as a medical device is provided. Recommendations addressing cytopathology‐specific benchmarks, multi‐institutional validation, transparent governance, and incremental deployment beginning with low‐risk applications are suggested. In the current environment, LLMs have the potential to augment cytopathology practice, but responsible adoption requires rigorous validation and sustained collaboration among cytopathologists, AI researchers, and regulatory bodies.

3118 Background: Improved risk stratification is critically needed across multiple cancers to guide treatment decisions. This includes reducing overtreatment in Ductal Carcinoma in Situ (DCIS), addressing late recurrence risk in estrogen receptor-positive/lymph node-negative (ER+/LN-) breast cancer (BC), and refining prognosis in head & neck squamous cell carcinomas (HNSCC), and high-grade serous ovarian carcinoma (HGSOC). Tumor-infiltrating lymphocytes (TILs) reflect anti-tumor immunity, whereas tumor-associated macrophages (TAMs) polarize toward anti-tumor M1-like (CD68+) or immunosuppressive M2-like (CD163+) phenotypes. Integrating TILs with TAM polarization markers may better capture inflammatory balance and improve outcome prediction. We developed a hematoxylin and eosin (H&E)-only AI method (VAIP) to quantify TILs and virtual CD68/CD163 TAMs to develop multi-cancer based potential prognostic models. Methods: TILs were detected on H&E with HoVer-UNet, and virtual CD68 and CD163 macrophages were inferred with VISTA. In cancer regions, we computed cell densities and immune-balance ratios and trained cancer-specific Cox models with compact feature sets (DCIS 6; HGSOC 7; ER+ 4; HNSCC 4). Prognostic stratification used disease-free survival (DFS) for DCIS (N=273), HGSOC (training N=165, independent test N=220), and ER+ BC (training N=144, test N=622), and overall survival (OS) for HNSCC (N=77), reflecting clinical endpoints. Evaluation used cross-validation for cohorts without an external test set (DCIS, HNSCC) and independent training and testing when independent cohort was available (HGSOC, ER+). Results: Risk stratification performance across cancer types is in Table 1. The DCIS model used 6 features, largely TIL-macrophage contrast ratios (TIL/M2, TIL/M1). HGSOC used 7 features with a similar ratio-heavy immune-balance signature, plus macrophage-normalized density and polarization metrics (e.g., M1/N, M/N). In contrast, ER+ BC and HNSCC achieved best performance with 4 features dominated by macrophage measures, particularly M2 density and the M1/M2 ratio. Across sites, CD163-related macrophage metrics and immune-balance ratios were consistently informative. Conclusions: Integrating H&E-visible TILs with virtual CD68/CD163 TAMs enables fully automated, H&E-only prognostic biomarkers that stratify OS and DFS across multiple cancers, demonstrating consistent prognostic value in DCIS, ER+ BC, HGSOC, and HNSCC. These findings support further study of this histomorphometric risk classifier through broader external and prospective validation. Performance across cancer types. Cancer Type Cohort C-index HR [95% CI] p-value DCIS All 0.616 1.76 [1.20-2.59] <0.001 DCIS RT/No-treatment 0.624 1.94 [1.15-3.26] 0.003 HGSOC Test 0.553 1.38 [1.02-1.87] 0.030 ER+ BC Test 0.577 1.51 [1.00-2.28] 0.004 HNSCC All 0.603 1.92 [1.03-3.36] 0.012

DEK::AFF2 fusion nonkeratinizing squamous cell carcinoma (NKSCC) is an emerging entity in the sinonasal tract, temporal bone, and skull base. However, the clinical behavior of these tumors has not been well studied. Here, we report the largest cohort of DEK::AFF2 carcinomas to determine if morphology, mitotic rate, and/or Ki-67 IHC are associated with patient outcomes, including a comparison with high-risk human papillomavirus (HPV)-associated and independent patients. We solicited cases of molecularly or AFF2 immunohistochemistry (IHC) proven DEK::AFF2 SCC from surgical pathologists to collect patient demographic, clinical, and outcome data. Using representative H&E slides, we characterized the morphology and counted mitoses. Ki-67 immunohistochemistry was performed. We also compared the DEK::AFF2 survival rates to those in a cohort of AFF2 IHC-negative HPV-associated and HPV-independent SCC. DEK::AFF2 carcinomas most commonly arose in the nasal cavity (13/30, 43%), and the average number of recurrences was 1.8 (range: 0 to 10). At the last follow-up, most patients were disease free (19/30, 63%) or were alive with disease (9/30, 30%). There was an average mitotic rate of 2 per 2 mm2 (range: 0 to 9) and Ki-67 proliferation rate of 26% (range: 3% to 60%). Local recurrence was common, but morphology, mitotic activity, and Ki-67 index were not associated with recurrence or survival. On Kaplan-Meier survival analysis, DEK::AFF2 patients had lower disease-free survival but otherwise had similar outcomes to conventional SCC patients. Our multi-institutional study shows that local recurrence is common in DEK::AFF2 fusion nonkeratinizing SCC patients, but patients have survival rates similar to conventional SCC. Despite showing a range of different features and proliferation rates, traditional grading by morphology, mitotic rate, and/or Ki-67 activity does not seem to be predictive of outcome.