JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

Development and evaluation of deep learning models for detecting and classifying various bone tumours in full-field limb radiographs using automated object detection models

  • 0Department of Orthopaedic Surgery, Graduate School of Medical Sciences, Kyushu University, Fukuoka, Japan.
Bone & Joint Research +

|

September 1, 2025

|

September 1, 2025

View abstract on PubMed

Summary

This summary is machine-generated.

A deep learning model called Detection with transformers with Improved deNoising anchOr boxes (DINO) shows improved detection of bone tumors compared to YOLO. DINO may help clinicians accurately detect and classify bone tumors in radiographs.

Area Of Science

  • Orthopaedic oncology
  • Radiology
  • Artificial intelligence in medicine

Background

  • Accurate detection and classification of bone tumors from radiographs are crucial for patient management.
  • Current diagnostic processes rely heavily on expert interpretation, which can be time-consuming and subject to inter-observer variability.

Purpose Of The Study

  • To develop and evaluate a fully automated deep learning model for detecting and classifying benign and malignant bone tumors in limb radiographs.
  • To compare the classification performance of the automated model against orthopaedic oncologists and general orthopaedic surgeons.

Main Methods

  • A retrospective analysis of 642 limb bone tumors from three institutions was performed.
  • End-to-end object detection models, Detection with transformers with Improved deNoising anchOr boxes (DINO) and You Only Look Once (YOLO), were employed.
  • Five-fold cross-validation was used for model training, parameter optimization, and performance evaluation.

Main Results

  • The DINO model achieved a higher mean tumor detection rate (85.7%) than the YOLO model (80.1%).
  • DINO demonstrated significantly higher accuracy and sensitivity compared to general orthopaedic surgeons.
  • The DINO model successfully classified challenging cases misclassified by multiple clinicians, though errors occurred with diagnostically difficult or unusually sited tumors.

Conclusions

  • The DINO model outperforms YOLO in automatically detecting bone tumors.
  • This deep learning approach shows potential to assist clinicians in the detection and classification of bone tumors in clinical practice.