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

Listen to Brian and Lisa discuss the paper 'Development and evaluation of deep learning models for detecting and classifying various bone tumours in full-field limb radiographs using automated object detection models' published in the September 2025 issue of Bone & Joint Research.

Click here to read the paper.

Be the first to know when the next episode is live! Follow our social media accounts, @BoneJointPortal and @BoneJointRes on X, and @bonejointres on Instagram for updates!

Transcript

[00:00:00] Welcome back to another episode of AI Talks with Bone & Joint, brought to you by the publishers of Bone & Joint Research. Today we're discussing the paper 'Development and evaluation of deep learning models for detecting and classifying various bone tumours in full-field limb radiographs using automated object detection models', published in September 2025 by M Yamana, and colleagues. I'm Brian and I'm here with my co-host Lisa.

Hello, Brian. This research is truly fascinating and could have a significant impact in the medical field. Where shall we begin?

Let's start with the motivation behind the research. The primary aim was to develop a fully automated deep learning model to detect and classify benign and malignant bone tumors and full-field limb radiographs using an object detection model.

Additionally, they aim to compare the performance of this model against three orthopaedic oncologists and three general orthopaedic surgeons.

How did they conduct the study? The study included 642 [00:01:00] limb bone tumors from three institutions with 378 being benign and 264 malignant. They used two object detection models, end-to-end object detection with transformers with Improved deNoising anchOr boxes known as DINO, and You Only Look Once or YOLO. They also performed five-fold cross validation on the collected radiographs.

Indeed, and one of the key findings was that the DINO model outperformed the YOLO model in terms of detection rate. DINO achieved an 85.7% mean tumor detection rate compared to YOLO's 80.1%. 

Exactly. They found that DINO was more accurate, especially in detecting malignant tumors. They also compared the classification performance of DINO with that of doctors using various metrics like accuracy, sensitivity, specificity, precision, and F-measure.

What impressed me was that the DINO model not only had a higher detection rate, but its classification, [00:02:00] accuracy, and sensitivity were also higher than those of the general orthopaedic surgeons.

For instance, the DINO model correctly classified 78.6% of the challenging cases that had been misclassified by two or more doctors. Indeed, this clearly shows the potential of the DINO model to assist doctors in clinical practice reducing the risk of misdiagnosis of bone malignancies.

A key point for me was the discussion on why the DINO model performed better.

The transformers architecture, which DINO leverages, uses an attention mechanism that evaluates the relationship among distant parts in the image making it particularly effective for detecting and classifying bone tumors.

Yes, that's an essential point. The YOLO model, based on a convolutional neural network, focuses more on local features and isn't as effective in capturing the broader context needed for bone tumor detection and classification.

It's also worth noting some limitations they acknowledged. The study had a [00:03:00] relatively small sample size and didn't include cases with non-tumor conditions like degenerative diseases. Plus the input image resolution was limited due to GPU memory constraints.

These are valid points. They did mention future studies could involve higher resolution images and broader data sets, potentially including other imaging methods for more comprehensive diagnoses.

In summary, the DINO model shows significant promise as a clinical decision support tool for detecting and classifying bone tumors. While there's still room for improvement, the transformer-based approach seems to offer a solid advantage over traditional CNN-based models.

Absolutely, it's a step forward in integrating AI into healthcare, potentially easing the workload on doctors and improving diagnostic accuracy. So what's the final takeaway for our listeners, Lisa?

The main takeaway is that artificial intelligence, particularly transformer-based models like DINO, holds great promise in the early detection and [00:04:00] classification of bone tumors. This could lead to better patient outcomes and more efficient clinical workflows.

Well said, that wraps up today's episode. Thanks for tuning into AI Talks with Bone & Joint. Until next time.