To design and test a patch-based deep learning model to classify H&E images of osteosarcoma into tumor vs. non-tumor and viable vs. non-viable, which is an alternative to the traditional necrosis assessment, for tumor screening and viable tissue stratification for treatment response. Eight hundred ninety-seven cleaned histopathology image patches (primary) and 1,048 public image patches (secondary) were analyzed. ResNet18, EfficientNet-B0, ViT-Tiny, and FastCNN were tested to classify tumor/non-tumor. ResNet18 was then used for multiclass viability classification and compared to a two-stage gated cascade. Confidence intervals, McNemar testing, ablation studies, duplicate and case-overlap audits, confusion matrices, ROC analysis, saliency confidence-drop evaluation, and Grad-CAM visualization were all used to evaluate performance. The framework achieved an accuracy of 97.8% and a support-adjusted macro-F1 score of 0.943 on Dataset 1. On Dataset 2, it achieved an accuracy of 96.0%, a macro-F1 score of 0.958, and a macro-AUC of 0.995, demonstrating strong reproducibility and discriminative performance. The proposed framework offers clinically interpretable and reproducible computational pathology support for the treatment-response assessment in osteosarcoma at the patch level. In the future, patient-level validation and whole-slide analysis, as well as prospective expert evaluation, should be performed.

To design and test a patch-based deep learning model to classify H&E images of osteosarcoma into tumor vs. non-tumor and viable vs. non-viable, which is an alternative to the traditional necrosis assessment, for tumor screening and viable tissue stratification for treatment response. Eight hundred ninety-seven cleaned histopathology image patches (primary) and 1,048 public image patches (secondary) were analyzed. ResNet18, EfficientNet-B0, ViT-Tiny, and FastCNN were tested to classify tumor/non-tumor. ResNet18 was then used for multiclass viability classification and compared to a two-stage gated cascade. Confidence intervals, McNemar testing, ablation studies, duplicate and case-overlap audits, confusion matrices, ROC analysis, saliency confidence-drop evaluation, and Grad-CAM visualization were all used to evaluate performance. The framework achieved an accuracy of 97.8% and a support-adjusted macro-F1 score of 0.943 on Dataset 1. On Dataset 2, it achieved an accuracy of 96.0%, a macro-F1 score of 0.958, and a macro-AUC of 0.995, demonstrating strong reproducibility and discriminative performance. The proposed framework offers clinically interpretable and reproducible computational pathology support for the treatment-response assessment in osteosarcoma at the patch level. In the future, patient-level validation and whole-slide analysis, as well as prospective expert evaluation, should be performed.