[KongchangAI]
ModelD2Mamba / D²Mamba

D²Mamba

A Mamba-type state space model for medical image segmentation that combines non-causal state space duality with deformable Mamba scanning, designed specifically for segmenting small retinal lesions such as microaneurysms.

Timeline (last 90 days)

Nov 11

The D²Mamba model achieved 33.87% mIoU in five-fold cross-validation on the DDR dataset, 2.99 percentage points higher than the strongest baseline.

Unverified50%
Nov 11

D²Mamba improved the Dice coefficient by 10.06 percentage points on the microaneurysm (MAs) segmentation subtask, the largest gain among all lesion types.

Unverified50%
Nov 11

The core designs of D²Mamba include non-causal state space duality and deformable Mamba scanning.

Unverified50%
Nov 11

Ablation experiments confirm that non-causal duality and deformable scanning each contribute to D²Mamba's performance, and that the two act synergistically.

Unverified50%
Nov 11

The code and pretrained weights of D²Mamba have been open-sourced in the GitHub repository https://github.com/PUREPANEL1/D2Mamba.

Unverified50%
Nov 11

This study has so far completed five-fold cross-validation on only a single dataset (DDR); cross-dataset generalization ability and stability under different imaging devices still require further testing.

Unverified50%

All Facts (6)

Source Articles