DiffGeo-AOR: Diffusion-Optimized Medical Grading via Geometric Priors enhanced Autoregressive Ordinal Regression

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Ordinal Regression

Background:

• Ordinal regression leverages category order but faces challenges in medical grading due to imbalanced data and ambiguous thresholds.
• Existing methods struggle with long-tailed distributions and inconsistent class boundaries in disease severity assessment.

Purpose of the Study:

• To propose DiffGeo-AOR, a novel autoregressive ordinal regression framework for medical grading tasks.
• To address data imbalance and threshold ambiguity issues inherent in medical severity classification.

Main Methods:

• DiffGeo-AOR employs an autoregressive process on continuous global features, decomposing K-class problems into K-1 binary decisions.
• Utilizes parameterized diffusion optimization for conditional probability modeling and FiLM-gated fusion for step-wise guidance.
• Incorporates rank-anchored ordinal priors for stable autoregressive module convergence.

Main Results:

• DiffGeo-AOR consistently outperforms state-of-the-art ordinal regression methods on 2D and 3D medical grading tasks.
• Demonstrated superior performance across three large-scale datasets, validating its effectiveness.
• The model successfully handles challenges of uneven data distribution and ambiguous severity thresholds.

Conclusions:

• DiffGeo-AOR offers a robust and effective solution for ordinal regression in medical image grading.
• The proposed method enhances classification accuracy by directly addressing data imbalance and threshold ambiguity.
• Publicly available code facilitates further research and application in medical AI.