Optimized data augmentation for osteosarcoma detection in deep and lightweight networks

Osteosarcoma (Ost) is an extremely aggressive primary bone malignancy that mostly occurs among children and young adults. Precise histopathological classification is challenging due to strong intra- and inter-tumoral heterogeneity, combined with the scarcity of annotated datasets. The current study demonstrates a systematic deep learning (DL) methodology crafted to investigate the effects of preprocessing and data augmentation approaches to osteosarcoma image classification. Hematoxylin and Eosin (H&E)-stained histopathological images were obtained from the publicly accessible UT Southwestern/UT Dallas Osteosarcoma dataset and standardized to standard noise reduction, contrast enhancement, and artifact suppression procedures to facilitate tissue prominence. Controlled augmentation settings were built (no augmentation, and 650, 1000, and 1500 synthetic images per class) to investigate how sequential dataset enlargement impacts generalization performance. Four transfer learning models similar to VGG19, InceptionV3, InceptionResNetV2, and NasMobileNet were fine-tuned and assessed through accuracy, sensitivity, specificity, and ROC-AUC metrics. The results confirm that moderate augmentation provided the best results, with NasMobileNet reporting 95.07 % accuracy, 95 % sensitivity, and 95 % specificity (AUC = 0.96), whereas deeper models like InceptionResNetV2 took advantage of increased augmentation (up to 94.37 % accuracy). Statistical analysis further confirmed that the found differences were not significant (p > 0.05), indicating support for consistency among models. The results overall highlight that the efficacy of augmentation depends on the model and that integration of systematic analysis with interpretability enhances the reliability of osteosarcoma classification through the power of deep learning.