Detecting Bone Lesions in X-Ray Under Diverse Acquisition Conditions

Tal Zimbalist, Ronnie Rosen, Keren Peri-Hanania, Yaron Caspi, Bar Rinott, Carmel Zeltser-Dekel, E. Bercovich, Yonina C. Eldar, Shai BagonJournal of Medical Imaging 2022

Abstract

The diagnosis of primary bone tumors is challenging, as the initial complaints are often non-specific. Early detection of bone cancer is crucial for a favorable prognosis. Incidentally, lesions may be found on radiographs obtained for other reasons. However, these early indications are often missed. In this work, we propose an automatic algorithm to detect bone lesions in conventional radiographs to facilitate early diagnosis. Detecting lesions in such radiographs is challenging: first, the prevalence of bone cancer is very low; any method must show high precision to avoid a prohibitive number of false alarms. Second, radiographs taken in health maintenance organizations (HMOs) or emergency departments (EDs) suffer from inherent diversity due to different X-ray machines, technicians and imaging protocols. This diversity poses a major challenge to any automatic analysis method. We propose to train an off-the-shelf object detection algorithm to detect lesions in radiographs. The novelty of our approach stems from a dedicated preprocessing stage that directly addresses the diversity of the data. The preprocessing consists of self-supervised region-of-interest detection using vision transformer (ViT), and a foreground-based histogram equalization for contrast enhancement to relevant regions only. We evaluate our method via a retrospective study that analyzes bone tumors on radiographs acquired from January 2003 to December 2018 under diverse acquisition protocols. Our method obtains 82.43% sensitivity at 1.5% false-positive rate and surpasses existing preprocessing methods. For lesion detection, our method achieves 82.5% accuracy and an IoU of 0.69. The proposed preprocessing method enables to effectively cope with the inherent diversity of radiographs acquired in HMOs and EDs.

Introduction

Primary bone tumors are rare but carry a high mortality rate, particularly among younger populations. Early detection is critical for a favorable prognosis, yet many lesions are discovered incidentally on radiographs taken for unrelated reasons, such as fractures or sprains. In these contexts, subtle indicators of bone cancer are frequently overlooked by clinicians focused on the primary injury.

Automating lesion detection is complicated by significant "data drift." Radiographs from various healthcare settings exhibit high diversity due to differences in X-ray equipment, technician settings, and imaging protocols. This variability, combined with the low prevalence of bone cancer, necessitates a detection system with both high sensitivity and high precision to avoid excessive false alarms.

Example of a clinical radiograph showing typical input diversity and background noise.

Methods

The proposed approach addresses data diversity through a dedicated preprocessing pipeline before applying a standard object detection model.

1. Self-Supervised ROI Detection

The system first isolates the bone structure (Region of Interest) from the background. This is achieved using DINO-ViT, a Vision Transformer trained in a self-supervised manner. The researchers found that the model's attention maps could accurately segment bone structures without requiring manual medical labeling, effectively transferring features learned from natural images to the medical domain.

2. Foreground-Aware Histogram Equalization

Standard contrast enhancement often fails on X-rays because the large areas of black background skew the global histogram. By using the masks generated in the first step, the system applies histogram equalization exclusively to the bone pixels. This standardizes the dynamic range and enhances relevant features regardless of the original acquisition conditions.

3. Lesion Detection

The preprocessed, high-contrast images are fed into a Faster R-CNN detector (specifically the Detectron2 implementation) to locate and classify potential lesions.

The complete pipeline: (a) raw input, (b) input histogram, (c) ROI mask, (d) enhanced image, (e) enhanced histogram, and (f) final lesion detection.

Results

The method was evaluated on a retrospective dataset of 1,421 radiographs spanning 15 years (2003–2018). The results demonstrate that addressing diversity at the input level significantly improves performance over standard enhancement techniques like CLAHE.

  • Sensitivity and Precision: The model achieved 82.43% sensitivity at a remarkably low 1.5% false-positive rate.
  • Detection Quality: For bounding box placement, the model reached 82.5% accuracy with an Intersection over Union (IoU) of 0.69.
  • Robustness: The system achieved an AUC of 0.98, outperforming baseline methods (0.94–0.95) and showing superior resilience to background artifacts such as clothing or medical equipment.

Conclusion

This work highlights that preprocessing is a critical component in deploying AI for medical imaging in real-world, diverse clinical environments. By leveraging the zero-shot capabilities of Vision Transformers for ROI detection and focusing contrast enhancement on the foreground, the proposed method effectively mitigates the challenges of data drift, facilitating the early detection of rare bone lesions.

Summary updated

Abstract

The diagnosis of primary bone tumors is challenging, as the initial complaints are often non-specific. Early detection of bone cancer is crucial for a favorable prognosis. Incidentally, lesions may be found on radiographs obtained for other reasons. However, these early indications are often missed. In this work, we propose an automatic algorithm to detect bone lesions in conventional radiographs to facilitate early diagnosis. Detecting lesions in such radiographs is challenging: first, the prevalence of bone cancer is very low; any method must show high precision to avoid a prohibitive number of false alarms. Second, radiographs taken in health maintenance organizations (HMOs) or emergency departments (EDs) suffer from inherent diversity due to different X-ray machines, technicians and imaging protocols. This diversity poses a major challenge to any automatic analysis method. We propose to train an off-the-shelf object detection algorithm to detect lesions in radiographs. The novelty of our approach stems from a dedicated preprocessing stage that directly addresses the diversity of the data. The preprocessing consists of self-supervised region-of-interest detection using vision transformer (ViT), and a foreground-based histogram equalization for contrast enhancement to relevant regions only. We evaluate our method via a retrospective study that analyzes bone tumors on radiographs acquired from January 2003 to December 2018 under diverse acquisition protocols. Our method obtains 82.43% sensitivity at 1.5% false-positive rate and surpasses existing preprocessing methods. For lesion detection, our method achieves 82.5% accuracy and an IoU of 0.69. The proposed preprocessing method enables to effectively cope with the inherent diversity of radiographs acquired in HMOs and EDs.
BibTeX
@article{zimbalist2022detecting,
  author = {Tal Zimbalist and Ronnie Rosen and Keren Peri-Hanania and Yaron Caspi and Bar Rinott and Carmel Zeltser-Dekel and E. Bercovich and Yonina C. Eldar and Shai Bagon},
  title = {{Detecting Bone Lesions in X-Ray Under Diverse Acquisition Conditions}},
  journal = {Journal of Medical Imaging},
  year = {2022},
}
Figure