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

Ordinal regression is well-known for leveraging the underlying inherent order between successive categories to obtain additional regularization beyond traditional probabilistic classification mechanism. However, there are challenges in real-world medical grading tasks: 1) The uneven distribution of disease severity levels, characterized by a long-tailed format, complicates the ordinal regression process. 2)The ambiguity in establishing disease severity thresholds introduces substantial challenges, rendering the ordinal regression framework susceptible to inter-class inconsistencies. To address the challenge, this work proposes DiffGeo-AOR, by introducing an autoregressive process to ordinal regression that operates directly on continuous global features, without any need for vector quantization. DiffGeo-AOR decomposes a $K$ -class ordinal problem into ${K}{-}{1}$ conditional binary decision steps, enabling the model to explicitly infer whether the severity has crossed the next grade threshold at each step. We also introduces parameterized diffusion optimization to model conditional probability distributions, allowing continuous global features to be extracted and directly leveraged in the autoregressive process. In addition, we design a FiLM-gated Step-Aware Diffusion Conditioning Fusion that guides each step's decision based on both the current image representation and the previous soft prediction probabilities. Furthermore, we regularize the feature space with rank-anchored ordinal priors during training to facilitate stable convergence of the autoregressive module. DiffGeo-AOR consistently outperforms current state-of-the-art ordinal regression methods across both 2D and 3D medical grading tasks on three large-scale datasets. The implementation code is publicly available at https://github.com/Qinkaiyu/DiffGeo-AOR.