Optimized data augmentation for osteosarcoma detection in deep and lightweight networks.

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Osteosarcoma (Ost) is an aggressive bone cancer primarily affecting young individuals.
• Histopathological classification of Ost is difficult due to tumor heterogeneity and limited annotated data.
• Deep learning (DL) offers potential for improving Ost classification accuracy.

Purpose of the Study:

• To systematically evaluate the impact of preprocessing and data augmentation on DL-based osteosarcoma image classification.
• To compare the performance of different transfer learning models (VGG19, InceptionV3, InceptionResNetV2, NasMobileNet) for Ost classification.
• To determine optimal augmentation strategies for enhancing DL model generalization in osteosarcoma diagnosis.

Main Methods:

• Utilized Hematoxylin and Eosin (H&E)-stained images from the UT Southwestern/UT Dallas Osteosarcoma dataset.
• Applied standardized preprocessing techniques including noise reduction and contrast enhancement.
• Implemented controlled data augmentation, generating 0, 650, 1000, and 1500 synthetic images per class.
• Fine-tuned four transfer learning models and evaluated using accuracy, sensitivity, specificity, and ROC-AUC.

Main Results:

• NasMobileNet achieved 95.07% accuracy, 95% sensitivity, and 95% specificity (AUC=0.96) with moderate augmentation.
• Deeper models like InceptionResNetV2 showed improved performance with increased augmentation, reaching 94.37% accuracy.
• Statistical analysis (p > 0.05) indicated no significant differences in performance across models, suggesting consistency.
• The effectiveness of data augmentation was found to be model-dependent.

Conclusions:

• Deep learning, combined with systematic analysis and interpretability, can enhance the reliability of osteosarcoma classification.
• The choice of data augmentation strategy should be tailored to the specific deep learning model used.
• This study provides a framework for optimizing DL methodologies for challenging histopathological image classification tasks in oncology.