Deep Learning and Radiomics Differentiate Benign vs Malignant Orbital Tumors on CT

Overview

This study evaluated deep learning-based radiomics (DL) and hand-crafted radiomics (HCR) on CT imaging to distinguish benign from malignant orbital tumors in 145 patients. The fused model combining DL and HCR features outperformed individual approaches, and a nomogram integrating clinical and semantic CT features showed superior predictive performance and clinical utility.

Background

Orbital tumors vary widely, with benign lesions often requiring observation and malignant tumors necessitating surgery. Accurate differentiation is critical but challenging due to overlapping clinical and imaging features. While histopathology is the gold standard, it requires invasive biopsy with associated risks. Advances in imaging and radiomics, including deep transfer learning, offer promising non-invasive diagnostic alternatives by extracting quantitative imaging features beyond human visual assessment.

Data Highlights

Key Findings

• Homogeneous enhancement and ill-defined/infiltrative margins on CT were independent semantic features differentiating benign from malignant tumors.
• A total of 14 hand-crafted radiomics and 30 deep learning features were extracted; 36 features were retained after fusion.
• The fused radiomics model achieved the highest AUC in the training cohort (0.986), outperforming single radiomics models.
• The nomogram combining clinical data, semantic CT features, and fused radiomics showed superior predictive performance and clinical utility, with test cohort AUC of 0.837.
• DeLong test showed no significant difference between the fused model and nomogram, but both outperformed individual models significantly.
• Decision curve analysis indicated the nomogram provided the highest net clinical benefit for differentiating orbital tumors.

Clinical Implications

Integrating deep learning and traditional radiomics features from routine contrast-enhanced CT can non-invasively and accurately differentiate benign from malignant orbital tumors. The nomogram incorporating clinical and semantic imaging features further enhances diagnostic confidence and may reduce the need for invasive biopsy procedures, especially in patients at risk for surgical complications.

Conclusion

The combined deep learning and hand-crafted radiomics approach, integrated into a clinically interpretable nomogram, offers a promising non-invasive tool for preoperative differentiation of orbital tumors. This method may improve clinical decision-making and patient management by accurately identifying tumor malignancy without invasive procedures.

References

1. Authors/Quzhou People’s Hospital/2024 -- Distinguishing Benign from Malignant Orbital Tumors through Deep Learning and Traditional Radiomics Analysis of CT Imaging