DFMA-ICH: a deformable mixed-attention model for intracranial hemorrhage lesion segmentation based on deep supervision

Intracranial hemorrhage (ICH) is a common and critical disease in clinical, with rapid progression, high disability, and mortality rates. Existing segmentation methods, such as U-Net and TransUNet, perform poorly for the problems in ICH segmentation such as small bleeding lesions, partial volume effects, and edge tissue adhesion with edema in ICH data. To solve these issues more efficiently, a deformable mixed-attention model based on deep supervision for ICH lesion segmentation (DFMA-ICH) is proposed in this study. DFMA-ICH consists of the short-term dense concatenate network (STDC) as the backbone with a mixed-attention method, an attention refining residual module (ARRM), and a mixed feature fusion module (MFFM). The mixed-attention method includes multi-scale spatial attention (MSP) and channel attention mechanism (SE) to extract rich lesion information. The double-pooling attention module (DPA) in ARRM is designed to correct features. In MFFM, different attention modules are constructed to reasonably combine low- and high-level features, and deformable convolution (DConv) is applied for boundary optimization. DFMA-ICH is trained by the deep supervision method to equalize the corresponding outputs at different stages. Overall, DFMA-ICH outperforms other advanced models on both spontaneous and traumatic ICH datasets by transfer learning with the Dice of 86.03, 80.98%, and HD of 12.35, 47.28 mm, respectively. Moreover, DFMA-ICH incurs the lowest time-space cost and exhibits the fastest inference speed. The study confirms that the proposed DFMA-ICH can provide an accurate and efficient method for ICH segmentation.