Radiograph-facilitated triage of primary bone tumors is critical for timely referral; however, it remains difficult in settings without subspecialty expertise. Building on the multicenter Bone Tumor X-ray Radiograph Dataset (BTXRD), in this study, we aimed to systematically benchmark modern deep learning architectures for classifying normal bone, benign tumors, and malignant tumors using a unified, reproducible pipeline. We evaluated a broad family of compound-scaled Convolutional Neural Networks (CNNs) and vision transformers using standardized preprocessing, class-imbalance-aware training via effective number weighting, and variance-aware k-fold cross-validation. Furthermore, we compared their performance with the BTXRD YOLOv8 Classification baseline. In addition to backbone evaluation, we conducted an architecture-aware analysis of contrast-limited adaptive histogram equalization (CLAHE), quantifying its effects on classification performance and learned representation geometry. To characterize representation shifts, we embedded penultimate-layer features using Uniform Manifold Approximation and Projection (UMAP) and determined how CLAHE alters neighborhood structure and class separability across backbones. We also integrated High Resolution Class Activation Mapping (HiResCAM) to assess explanation faithfulness and elucidated the degree of spatial overlap between model attention and radiologist-provided lesion masks. The experimental protocol incorporated a frozen held-out test set, calibration-sensitive metrics, and per-image prediction auditing to reflect deployment-oriented behavior. Across architectures, EfficientNet-B6 with CLAHE was the most reliable configuration, delivering high accuracy and macro-F1, while maintaining clinically meaningful error patterns and precise malignant predictions. This work provides novel insights into how preprocessing, architectures, feature representations, and explanation methods interact in bone tumor X-ray classification and establishes reproducible baselines to support future research on BTXRD.