3D Segmentation of Intracerebral Hemorrhage in Brain CT Using Enhanced UNet Transformer via Reinforcement Learning

Sana Ullah, Kim Gabin, Dongwook Lee, Hafiz Zahid Tufail, Kangsan Lee, Sunghyun Park, Joon-Young Kim, Won Hee Lee, Tae-Seong Kim

Keywords: Intracranial Hemorrhage, 3D Segmentation, PPO, Reinforcement Learning

Registration Requirement: Yes

Abstract: Intracranial hemorrhage (ICH) segmentation from 3D CT is critical for treatment planning, as ICH can lead to severe neurological deficits and death. While 2D slice-wise convolutional neural networks and transformer-based models have achieved strong performance, 2D approaches may lose inter-slice continuity that is important for volumetric lesions. In contrast, 3D models capture spatial relationships across the volume but still struggle with accurate delineation of small or low-contrast hemorrhages. To address these limitations, we propose a novel segmentation framework enhanced by reinforcement learning. Our approach, named UNETR-PPO, is an integration of the transformer-based UNETR segmentation model with reinforcement learning. We evaluate our method on the HemSeg-200 3D CT ICH dataset and demonstrate improved performance over the baseline models including 3D U-Net, V-Net, SwinUNETR, and UNETR in terms of Dice, IoU, and HD95.

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 16