Automated Detection and Classification of Vertebral Compression Fractures in CT Images Using Deep Learning and Radiomics Features

Project Abstract

This study presents a robust artificial intelligence framework that automates the detection and classification of vertebral compression fractures (VCFs) in computed tomography (CT) images by integrating deep learning with radiomics features. The proposed pipeline begins with an automated vertebral segmentation module using a 3D U-Net architecture to delineate vertebral bodies from chest, abdomen, and spine CT scans, followed by a vertebral level labeling and quality assurance stage to ensure accurate localization. A multi-stage fracture detection engine then processes the extracted vertebral volumes through a convolutional neural network (CNN) configured to distinguish fractured from non-fractured vertebrae, leveraging transfer learning from pretrained medical imaging models to address limited labeled data. To enhance diagnostic performance and interpretability, radiomic features—encompassing first-order statistics, texture matrices (GLCM, GLRLM, GLSZM, GLDM), and shape descriptors—are computed from the segmented vertebral regions and fused with learned deep features via a late fusion strategy. Feature selection is employed using a combination of recursive feature elimination and LASSO regularization to reduce redundancy and optimize classifier input. The final classification is performed using an ensemble method that integrates a gradient boosting classifier with a calibrated probability estimator to provide reliable fracture risk scores and confidence intervals. The study introduces a novel labeling scheme for VCF severity into multiple clinically relevant categories (mild, moderate, severe) to support prognosis and treatment planning. A comprehensive dataset compiled from multi-institutional CT repositories, including sagittal and coronal reconstructions with varying slice thickness and field-of-view, is annotated by experienced radiologists to establish a robust ground truth. Data harmonization techniques are applied to mitigate cross-scanner variability, and external validation on an independent cohort demonstrates the model’s generalizability. The performance is evaluated against conventional radiologist assessment and existing automated methods using metrics such as accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, Cohen’s kappa for inter-rater agreement, and calibration curves for probability estimates. Explainability is addressed through saliency mapping and saliency-guided radiomics visualization to identify imaging regions driving predictions and to correlate radiomic signatures with fracture patterns. The integration of radiomics with deep learning is shown to improve fracture detection sensitivity by capturing subtle textural and morphologic cues that may be overlooked by pure deep learning approaches, especially in early-stage fractures. The framework supports clinical workflow by producing per-vertebra fracture risk scores, severity classification, and visual explanations that can assist radiologists in prioritizing reads and standardizing reporting. Limitations related to dataset heterogeneity, annotation variance, and potential biases are addressed with cross-validation, stratified sampling, and rigorous statistical testing. The work provides a step toward automated, explainable, and scalable VCF assessment in CT imaging, with potential extensions to longitudinal monitoring and post-treatment evaluation.

Project Overview

What This Project Is About

---

A simple study that looks at how CT scans can be used to detect if a vertebra is fractured, and how to tell what type it is, using computer-based methods. The project combines image analysis with pattern learning to help doctors assess spine injuries faster and more consistently.

---

---

The Problem It Addresses

---

---

---

Objectives of the Project

---

1. Learn how CT images show vertebral fractures and what features doctors use.
2. Build a simple model to automatically detect fractures in CT scans.
3. Classify fractures into common types with explanations that are easy to understand.
4. Evaluate the model against expert opinions to check accuracy.
5. Discuss how such a tool could help clinicians in practice.

---

---

What You Will Do Step by Step

---

1. Collect a small dataset of labeled CT images with vertebral fractures.
2. Preprocess images to clean noise and standardize formats.
3. Extract simple, meaningful features from the images that a computer can learn from.
4. Train a basic learning model to detect and classify fractures.
5. Test the model and compare its results with expert labels.
6. Interpret results in plain terms and note limitations.
7. Document steps and create a user-friendly explanation of the tool.
8. Suggest future improvements and potential clinical use cases.

---

---

Expected Outcome