Adapted Deep Ensemble Learning-Based Voting Classifier for Osteosarcoma Cancer Classification.

Area of Science:

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

Background:

• Osteosarcoma classification relies on histopathological analysis of H&E-stained images.
• Unevenly dispersed datasets pose challenges for developing reliable machine learning models.
• Data augmentation is crucial for enhancing model generalization.

Purpose of the Study:

• To develop and evaluate deep learning models for accurate osteosarcoma classification.
• To address data imbalance issues in osteosarcoma image datasets.
• To improve the reliability and performance of automated diagnostic tools.

Main Methods:

• Utilized a dataset of hematoxylin and eosin-stained osteosarcoma images.
• Employed data augmentation techniques to improve generalization.
• Developed a novel Convolutional Neural Network (CNN) and an ensemble voting classifier.
• Evaluated six pre-trained CNN models (MobileNetV1, MobileNetV2, ResNetV250, InceptionV2, EfficientNetV2B0, NasNetMobile) in frozen and fine-tuned phases.

Main Results:

• The proposed CNN model achieved a Kappa score of 93.09%, outperforming pre-trained models.
• The adapted heterogeneous ensemble-learning-based voting classifier attained the highest Kappa score of 96.50%.
• The ensemble model demonstrated superior performance compared to all other evaluated models.

Conclusions:

• The proposed deep learning models, particularly the ensemble voting classifier, show significant potential for accurate osteosarcoma classification.
• These findings have practical implications for telemedicine, mobile healthcare, and supporting medical professionals in diagnosis.
• Addressing data imbalance with an evenly distributed training dataset is key to developing unbiased learning models.