Ordinal Regression Research Based on Dual Loss Function-An Example on Lumbar Vertebra Classification in CT Images.

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Ordinal classification problems, common in medical contexts (e.g., low-medium-high severity), require specialized modeling.
• Accurate lumbar vertebra identification is crucial for clinical assessments like sarcopenia detection.
• Existing classification methods may not fully leverage the inherent order of categorical data.

Purpose of the Study:

• To propose and evaluate a dual loss framework, Ordinal Residual Dual Loss, for ordinal regression in lumbar vertebra classification.
• To enhance the accuracy of L3 identification and sarcopenia detection using deep learning on CT images.
• To investigate the benefits of incorporating ordinal information into classification loss functions.

Main Methods:

• A retrospective study utilizing lumbar spine CT images.
• Development of a deep learning model based on the MobileNet-v3-Large architecture.
• Implementation of a dual loss framework combining standard cross-entropy loss with a novel Ordinal Residual Loss.

Main Results:

• The Ordinal Residual Dual Loss framework demonstrated superior performance compared to using cross-entropy loss alone.
• The proposed method surpassed the performance of previous studies in lumbar vertebra classification.
• Experimental results confirm the effectiveness of the dual loss approach for ordinal classification tasks.

Conclusions:

• The dual loss design effectively incorporates ordinal information, significantly improving vertebral level classification on CT images.
• This approach shows potential for more accurate automated L3 localization and sarcopenia assessment in clinical practice.
• Ordinal regression techniques offer a promising direction for advancing medical image analysis.