Clinical Scorecard: Distinguishing Benign from Malignant Orbital Tumors through Deep Learning and Traditional Radiomics Analysis of CT Imaging

At a Glance

Key Highlights

• Multivariate analysis identified homogeneous enhancement and ill-defined/infiltrative margins on CT as independent features differentiating benign from malignant orbital tumors.
• The fused model combining deep learning and hand-crafted radiomics features achieved superior diagnostic accuracy (AUC up to 0.986 in training cohort) compared to single radiomics approaches.
• A nomogram integrating clinical data and semantic CT features demonstrated high predictive performance and clinical utility, supporting non-invasive diagnosis.

Guideline-Based Recommendations

Diagnosis

• Use contrast-enhanced CT imaging to assess orbital tumors with focus on lesion location, margin definition, enhancement pattern, calcification, necrosis, and internal density homogeneity.
• Apply deep transfer learning to extract deep learning radiomics features and traditional methods for hand-crafted radiomics features from CT images.
• Combine radiomics features with clinical and semantic CT features in a nomogram to improve differentiation between benign and malignant tumors.

Management

• Consider non-invasive radiomics-based diagnostic tools to support clinical decision-making, especially for patients unable to undergo invasive biopsy procedures.
• Use radiomics analysis to guide treatment planning by distinguishing benign lesions (often managed conservatively) from malignant tumors (requiring surgical intervention).

Monitoring & Follow-up

• Monitor lesion characteristics on follow-up CT imaging using radiomics features to assess changes suggestive of malignancy or treatment response.

Risks

• Recognize that invasive tissue sampling carries risks including surgical complications, functional impairment, disfigurement, and vision loss.
• Utilize non-invasive radiomics approaches to minimize biopsy-related risks where appropriate.

Patient & Prescribing Data

145 patients with pathologically confirmed orbital tumors (48 benign, 97 malignant) undergoing contrast-enhanced CT prior to surgery

Radiomics models, especially fused deep learning and hand-crafted features combined with clinical data, provide accurate non-invasive differentiation aiding treatment decisions without immediate biopsy.

Clinical Best Practices

• Perform contrast-enhanced CT within 2 weeks prior to surgery for optimal imaging quality.
• Ensure high-quality image acquisition minimizing motion or susceptibility artifacts to enable accurate radiomics feature extraction.
• Use multivariate analysis to identify key CT semantic features (homogeneous enhancement, margin definition) for integration with radiomics data.
• Apply LASSO regression for feature selection and model construction to optimize diagnostic performance.
• Incorporate nomogram tools combining radiomics and clinical data for individualized risk assessment and clinical decision support.

References

• Quzhou People’s Hospital Medical Ethics Committee Approval