Autoencoder-Assisted Stacked Ensemble Learning for Lymphoma Subtype Classification: A Hybrid Deep Learning and...

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Accurate lymphoma subtype identification is critical for effective cancer diagnosis and treatment.
• Standard deep learning methods face challenges with overfitting and limited generalization.
• There is a need for more robust and reliable methods for lymphoma classification.

Purpose of the Study:

• To develop an autoencoder-augmented stacked ensemble learning (SEL) framework for improved lymphoma subtype identification.
• To integrate deep feature extraction (DFE) with machine learning classifiers for enhanced diagnostic accuracy.
• To overcome limitations of traditional deep learning in lymphoma classification.

Main Methods:

• Utilized Convolutional Autoencoder (CAE) for high-level feature extraction from histopathological images.
• Applied Principal Component Analysis (PCA) for dimensionality reduction of extracted features.
• Employed an SEL approach with Gradient Boosting Machine (GBM) meta-classifier, integrating Random Forest (RF), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), AdaBoost, and Extra Trees classifiers.

Main Results:

• The stacked ensemble classifier achieved 99.04% accuracy, 0.9998 AUC, and 0.9996 AP, outperforming individual models and standard deep learning methods.
• Multi-Layer Perceptron (MLP) and Random Forest (RF) showed strong standalone performance.
• PCA and t-SNE visualizations confirmed effective class discrimination through DFE.

Conclusions:

• The autoencoder-assisted ensemble learning approach provides a highly accurate and reliable method for lymphoma classification.
• AI models offer interpretable outputs, assisting pathologists in validating diagnostic predictions.
• Future research should focus on computational efficiency and multi-center validation for broader generalizability.