3588 - Machine Learning Using >300,000 H&E Images Refines Cancer of Unknown Primary Subtyping
Presenter(s)
K. M. Boehm1, M. Darmofal1, A. Aukerman1, A. Pasha1, A. Kohli1, R. Lim1, T. Pollard1, D. Moore1, A. Begum1, N. Rekhtman1, H. Al-Ahmadie1, J. Chang1, D. R. Gomez2, H. Nagar1, J. Janopaul-Naylor1, L. Z. Braunstein1, J. Jee1, N. Schultz3, S. P. Shah1, and F. Sanchez-Vega1; 1Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 3Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY
Purpose/Objective(s): Optimal treatment of de novo metastatic cancers hinges upon identifying the primary site. Using clinical sequencing panels, recent works have identified subtypes of cancers of unknown primary (CUP) with greater granularity, adding prognostic and predictive value. However, clinical sequencing is not ubiquitous and can delay treatment. Hence, we examined the value of machine learning to infer histologic subtype from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) of tumor specimens.
Materials/Methods: We developed a machine learning model, AEON (adaptive embedding ontologic network), to infer tumor histologic subtype from H&E WSIs. We assembled a pan-cancer cohort of 315,383 H&E WSIs from 62,266 patients spanning 193 subtypes, the largest cohort of its kind. Patients were allocated to train, validation, and test sets using a 70/10/20% split. Using these data, we trained a transformer to infer OncoTree codes from WSIs.
Results: AEON attained a weighted macro average area under the receiver operating characteristic curve (AUROC) of 0.95 across histologic subtypes in the test set. AUROC was 0.98 [0.98 - 0.99] for lung adenocarcinoma (1287 test examples), 0.98 [0.97 - 0.98] for invasive ductal carcinoma (846 test examples), and 0.99 [0.99-0.99] for prostate adenocarcinoma (772 test examples). The model performed well for rare subtypes such as epithelioid hemangioendothelioma (8 test examples; 0.99 [0.99 - 1.0]) and ovarian granulosa cell tumor (GRCT; 13 test examples; 0.99 [0.99 - 1.0]). 82% of AEON-labeled GRCT samples exhibited C134W FOXL2 mutations. Brackets denote 95% confidence interval.
AEON largely refined the classification of CUP specimens. 292 test specimens were labeled by pathologists as CUP, adenocarcinoma (AD) not otherwise specified (NOS), squamous cell carcinoma (SCC) NOS, or undifferentiated malignant neoplasm. AEON reassigned 90 unique subtypes, most commonly mixed lobular and ductal carcinoma (N = 31). Molecular profiling corroborated AEON-assigned subtypes: e.g., samples relabeled cutaneous SCC (N = 7) and pancreatic adenocarcinoma (N = 11) were enriched for TERT (57% vs 0%) and KRAS variants (0% vs 64%), respectively. We next examined the model’s prognostic value by comparing OS across AEON-assigned subtypes. For the 15 subtypes with = 5 test examples, OS differed significantly (log-rank p = 0.045). Cases assigned as SCCNOS (N = 7), pleomorphic lung carcinoma (N = 28), cholangiocarcinoma (N = 14), and ADNOS (N = 11) exhibited median OS of 30, 16, 12, and 6m respectively, corroborating expected trends. Median follow up was 48m.Conclusion: AEON-reassigned subtypes prognosticate CUP cases, supporting the role of machine learning on H&E images to improve CUP subtyping. This extends prior tools to cases where only diagnostic H&E images are available, reducing time to initial therapy. AEON could also help identify primary sites in patients with concurrent primary cancers.