AI-Based Histopathology Demonstrates High Accuracy for FGFR3-TACC3 Fusion Classification in Non-Small Cell Lung Cancer
Clinical Summary:
-
Design/Population: This retrospective, multisite computational pathology study evaluated a gated attention multiple instance learning model for classifying FGFR3::TACC3 fusion-positive non-small cell lung cancer (NSCLC) using hematoxylin and eosin-stained whole-slide images.
-
Key Outcomes: The model achieved area under the receiver operating characteristic curve values ranging from 0.8846 to 0.9998 across 3 cohort configurations, with classification accuracy between 98.4% and 98.8%. Attention mapping identified image regions contributing to model predictions.
-
Clinical Relevance: These findings suggest that deep learning may detect morphologic patterns associated with FGFR3::TACC3 fusions in NSCLC, supporting further investigation of AI-assisted histopathology as a potential complementary approach to molecular profiling.
Results from a multisite computational pathology study demonstrated that a deep learning model accurately classified FGFR3::TACC3 fusion-positive non-small cell lung cancer (NSCLC) using hematoxylin and eosin-stained whole-slide images, supporting the potential of artificial intelligence to identify morphologic features associated with oncogenic gene rearrangements.
These results were presented by Anusha Kanagala, MS, University of Alabama at Birmingham, at the Association for Molecular Pathology (AMP) Annual Meeting and Expo.
In this study, researchers developed a supervised deep learning model using hematoxylin and eosin-stained whole-slide images from patients with NSCLC, including FGFR3::TACC3 fusion-positive cases obtained from a collaborating institution and control samples from The Cancer Genome Atlas repository and institutional archives.
The model was designed to identify histologic patterns associated with FGFR3::TACC3 fusions, which represent a subset of FGFR rearrangements that may predict sensitivity to FGFR-targeted therapies. Whole-slide images underwent automated tissue segmentation using Otsu thresholding, with artifact exclusion through connected-component filtering.
Investigators extracted up to 2000 nonoverlapping image patches per slide at 20× magnification, requiring more than 50% tissue coverage. A pretrained Swin-Tiny model generated 768-dimensional patch embeddings, while ComBat batch correction was applied to reduce technical variation across institutions. A gated attention multiple instance learning classifier was used to identify diagnostically relevant image regions without requiring patch-level annotations.
Model performance was evaluated using leave-one-out cross-validation across 3 cohort configurations.
In the mixed cohort without deduplication, which included 112 FGFR3::TACC3 fusion-positive samples and 50 controls, the model achieved an area under the receiver operating characteristic curve of 0.9998 and classification accuracy of 98.8%.
When the analysis was restricted to institutional controls, including 112 fusion-positive samples and 19 controls, the model achieved an area under the receiver operating characteristic curve of 0.8846 and accuracy of 98.4%.
In a separate analysis selecting the single representative slide with the greatest tissue coverage per case, the model achieved an area under the receiver operating characteristic curve of 0.9997 and accuracy of 98.4% among 66 fusion-positive cases and 58 controls.
Across the 3 cohort configurations, area under the receiver operating characteristic curve values ranged from 0.8846 to 0.9998, demonstrating the model's ability to distinguish fusion-positive from control whole-slide images.
Attention mapping provided additional insight into image regions influencing model predictions, suggesting that FGFR3::TACC3 fusions may be associated with subtle histologic patterns detectable through deep learning.
These findings support the potential use of artificial intellegence-based histopathology to identify morphologic features associated with FGFR3::TACC3 fusions, although the approach has not been established as a replacement for molecular testing.
Source:
Kanagala A, Bodduluri S, Garcia-Muniz J, et al. Deep learning-based classification of NSCLC with FGFR3-TACC3 fusion using whole slide images. Presented at the AMP Annual Meeting and Expo. November 10-14, 2026; Seattle, Washington. Abstract I065.


