Deep Learning for Primary Bone Tumor Analysis on Radiographs: A Narrative Review of Classification, Detection, Localization, and Explainability

Abstract

Primary bone tumors are uncommon but clinically important lesions with heterogeneous radiographic appearances, making accurate diagnosis challenging. Conventional radiography remains an important first-line imaging modality, while artificial intelligence (AI), particularly deep learning (DL), has increasingly been investigated for computer-aided analysis of musculoskeletal abnormalities. This narrative review synthesizes recent applications of DL for primary bone tumor analysis on conventional radiographs, with emphasis on classification, detection and localization, segmentation, multitask learning, transformer-based approaches, and explainable artificial intelligence (XAI). The reviewed studies are compared in terms of model architectures, datasets, task definitions, performance measures, external validation, comparison with human readers, and explainability strategies. The literature indicates that DL models can achieve promising performance in selected bone tumor classification and detection tasks, with some studies reporting results comparable to experienced radiologists. However, direct comparison across studies remains difficult because of differences in datasets, tumor categories, imaging settings, evaluation protocols, and validation strategies. Important challenges include the rarity of primary bone tumors, class imbalance, limited multicenter and external validation, heterogeneous radiographic data, and inconsistent evaluation of model explanations. Recent developments in publicly available datasets, lesion localization, transformer-based architectures, multimodal learning, and XAI provide opportunities for more robust and clinically relevant systems. Future research should prioritize larger multicenter datasets, patient-level validation, clinically meaningful lesion localization and explainability, integration of complementary clinical information, and prospective evaluation of AI-assisted diagnostic workflows.