Boosting predictive accuracy in tumor cellularity evaluation with AI-powered ensemble methods

Abstract

Purpose: Tumor cellularity Evaluation quantifies the proportion of tumor cells in a sample, providing a precise measure of tumor burden. However, cells in Hematoxylin and Eosin (H&E) images exhibit distinct features, arising from variability and overlap in appearance, making their distinguishing more challenging in real-world scenarios. This study aims to apply a homogeneous ensemble method on deep learning models to estimate TC values, thereby eliminating the labor-intensive and time-consuming aspects of the task. Methods: A feature ensemble model integrating multiple deep learning models trained on distinct feature sets was developed. This study employed homogeneous ensemble learning strategies and Test-time augmentation (TTA) to enhance prediction accuracy and robustness. Using Intraclass Correlation Coefficient (ICC) and Mean Absolute Error (MAE) metrics, comparisons reveal superior performance of ensembles over single models. Results: The feature ensemble model achieved significant improvements in accuracy, achieving an ICC of 0.981 on an internal testing set and 0.9437 on an external testing set. The comparison results demonstrated the superior performance of the ensemble approach, highlighting the improved accuracy. Conclusion: This study underscores the importance of ensemble methods in improving predictive accuracy and robustness in digital pathology. Amalgamating digital image analysis (DIA) with artificial intelligence (AI) exemplifies a breakthrough in tumor cellularity estimation. Employing a homogeneous ensemble learning strategy, the model not only excels in performance but also provides a holistic solution, surpassing the constraints of conventional single-model approaches. This advancement signifies a pivotal step in enhancing medical imaging analytics.