Exponentially weighted ensemble deep learning for histologic growth pattern classification in lung adenocarcinoma
- 1Roux Institute, Northeastern University
Abstract
Accurate classification of lung adenocarcinoma growth patterns from hematoxylin and eosin (H&E)-stained whole-slide images is essential for treatment planning and prognosis. We develop and evaluate an exponentially weighted ensemble deep learning approach that combines five architectures (EfficientNet-B3, DeiT3, Swin Transformer, Vision Transformer, ConvNeXt) with weights determined by a softmax transformation of per-model Quadratic Weighted Kappa (QWK) scores. Using 25,545 image patches at 40x magnification from the Dartmouth-Hitchcock dataset, we evaluated performance with stratified 5-fold cross-validation. Individual model performance ranged from Cohens kappa{kappa} = 0.5563 (Con-vNeXt) to{kappa} = 0.9561 (EfficientNet-B3). The ensemble achieved 96.52% accuracy and{kappa} = 0.9648, exceeding the best individual model by {Delta}{kappa} = +0.0087 with statistical significance (McNemars test,{chi} 2 = 50.84, p < 10-12; bootstrap 95% CI on {Delta}{kappa} [+0.0056, +0.0117]). Pairwise disagreement rates of 16.6-43.2% between architectures indicate that the models capture different aspects of tissue morphology, and the framework automatically down-weights unreliable components, as illustrated by ConvNeXt receiving only 5.7% ensemble weight despite a single-fold training failure. The approach offers a clinically grounded method for ensemble aggregation in digital pathology.
Related
- Case report and literature review: histologic transformation from EGFR-mutant lung adenocarcinoma to squamous cell carcinoma with newly detected SMAD4 and ATM variants
- Imaging of benign salivary gland tumors under the fifth WHO classification, with a focus on rare histologic types and newly listed entities—Part II
- Imaging of benign salivary gland tumors: an update based on the fifth WHO classification, with a focus on rare histologic types and newly listed entities—part I
- Non-Invasive Assessment of SMAD4 Expression Status in Pancreatic Ductal Adenocarcinoma: Diagnostic Value of T1ρ-Weighted Imaging Combined with Clinical Parameters
- Multi-Platform Transcriptomic Classification of LUAD Driver Mutation Subtypes via Heterogeneous Ensemble Learning
BibTeX
@misc{korutla2026exponentially,
title = {{Exponentially weighted ensemble deep learning for histologic growth pattern classification in lung adenocarcinoma}},
author = {R. Korutla and M. N. Jain and T. M. Karkera and G. Tunik and S. T. Ryan and R. A. Christman and S. Amal},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.24.26363902},
url = {https://www.medrxiv.org/content/10.64898/2026.09.24.26363902v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Exponentially weighted ensemble deep learning for histologic growth pattern classification in lung adenocarcinoma
AU - R. Korutla
AU - M. N. Jain
AU - T. M. Karkera
AU - G. Tunik
AU - S. T. Ryan
AU - R. A. Christman
AU - S. Amal
PY - 2026
DA - 2026/09/24
PB - medRxiv
DO - 10.64898/2026.09.24.26363902
UR - https://www.medrxiv.org/content/10.64898/2026.09.24.26363902v1
AB - Accurate classification of lung adenocarcinoma growth patterns from hematoxylin and eosin (H&E)-stained whole-slide images is essential for treatment planning and prognosis. We develop and evaluate an exponentially weighted ensemble deep learning approach that combines five architectures (EfficientNet-B3, DeiT3, Swin Transformer, Vision Transformer, ConvNeXt) with weights determined by a softmax transformation of per-model Quadratic Weighted Kappa (QWK) scores. Using 25,545 image patches at 40x magnification from the Dartmouth-Hitchcock dataset, we evaluated performance with stratified 5-fold cross-validation. Individual model performance ranged from Cohens kappa{kappa} = 0.5563 (Con-vNeXt) to{kappa} = 0.9561 (EfficientNet-B3). The ensemble achieved 96.52% accuracy and{kappa} = 0.9648, exceeding the best individual model by {Delta}{kappa} = +0.0087 with statistical significance (McNemars test,{chi} 2 = 50.84, p < 10-12; bootstrap 95% CI on {Delta}{kappa} [+0.0056, +0.0117]). Pairwise disagreement rates of 16.6-43.2% between architectures indicate that the models capture different aspects of tissue morphology, and the framework automatically down-weights unreliable components, as illustrated by ConvNeXt receiving only 5.7% ensemble weight despite a single-fold training failure. The approach offers a clinically grounded method for ensemble aggregation in digital pathology.
ER -