Qualitative Histopathological Classification of Primary Bone Tumors Using Deep Learning: A Pilot Study.

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Histopathological diagnosis of bone tumors presents significant challenges for pathologists.
• Accurate classification of bone tumor aggressiveness is crucial for effective treatment planning.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for classifying bone tumors based on histopathological aggressiveness.
• To compare the diagnostic performance of DL models against human pathologists.

Main Methods:

• Whole slide imaging (WSI) of 427 bone tumor slides were processed into 716,838 image patches.
• Six DL models were trained and validated for binary (benign vs. non-benign) and ternary (benign vs. intermediate vs. malignant) classification.
• Performance was assessed at patch and slide levels, with comparisons made against four pathologists using Cohen's kappa score (CKS).

Main Results:

• VGG-16 and Inception V3 models demonstrated superior performance in patch-level classification.
• For slide-level prediction, DL models achieved high accuracy (AUC 0.962-0.971 for binary, CKS 0.731-0.802 for ternary).
• DL model performance was comparable to senior pathologists and superior to junior pathologists, with visualization confirming reliance on pathological features.

Conclusions:

• Deep learning models can effectively classify bone tumor aggressiveness histopathologically.
• DL-assisted diagnosis shows performance on par with experienced pathologists, offering potential to improve diagnostic efficiency.
• These findings support the future integration of DL into clinical practice for bone tumor histopathology.