Transferable CNN-based data mining approaches for medical imaging: application to spine DXA scans for osteoporosis detection

TL;DR

A new CNN model is developed to detect osteoporosis in spine DXA scans, outperforming existing models with high accuracy.

Contribution

A novel 21-layer CNN model for osteoporosis detection in spine DXA scans is proposed and validated.

Findings

The proposed CNN model achieved a 97.16% F1-score and 97.14% accuracy in detecting osteoporosis.

It outperformed pre-trained models like ResNet-50, VGG-16, and InceptionV3.

The model's high precision (99.04%) and recall (95.41%) indicate strong diagnostic performance.

Abstract

Osteoporosis is the leading cause of sudden bone fractures. This is a silent and deadly disease that can affect any part of the body, such as the spine, hips, and knee bones.

To measure bone mineral density, dual-energy X-ray absorptiometry (DXA) scans help radiologists and other medical professionals identify early signs of osteoporosis in the spine.

A proposed 21-layer convolutional neural network (CNN) model is implemented and validated to automatically detect osteoporosis in spine DXA images. The dataset contains 174 spine DXA images, including 114 affected by osteoporosis and the rest normal or non-fractured. To improve training, the dataset is expanded using various data augmentation techniques.

The classification performance of the proposed model is compared with that of four popular pre-trained models: ResNet-50, Visual Geometry Group 16 (VGG-16), VGG-19, and InceptionV3. With an F1-score of 97.16%, recall of 95.41%, classification accuracy of 97.14%, and precision of 99.04%, the proposed model consistently outperforms competing approaches.

The proposed paradigm would therefore be very valuable to radiologists and other medical professionals. The proposed approach’s capacity to detect, monitor, and diagnose osteoporosis may reduce the risk of developing the condition.

Figures