EMIT-Diff: Enhancing Medical Image Segmentation via Text-Guided Diffusion Model

• DOI: 10.48550/arxiv.2310.12868
• Publication Date: 2023-01-01
• Authors: Zheyuan Zhang; Lanhong Yao; Bin Wang; Debesh Jha; Elif Keleş; Alpay Medetalibey...; Ulaş Bağcı
• Keywords: Overfitting; Leverage (statistics)

Large-scale, big-variant, and high-quality data are crucial for developing robust successful deep-learning models medical applications since they potentially enable better generalization performance avoid overfitting. However, the scarcity of labeled always presents significant challenges. This paper proposes a novel approach to address this challenge by controllable diffusion image synthesis, called EMIT-Diff. We leverage recent probabilistic generate realistic diverse synthetic that preserve essential characteristics original images incorporating edge information objects guide synthesis process. In our approach, we ensure synthesized samples adhere medically relevant constraints underlying structure imaging data. Due random sampling process model, can an arbitrary number with appearances. To validate effectiveness proposed method, conduct extensive set segmentation experiments on multiple datasets, including Ultrasound breast (+13.87%), CT spleen (+0.38%), MRI prostate (+7.78%), achieving improvements over baseline methods. For first time, best knowledge, promising results demonstrate EMIT-Diff tasks show feasibility introducing first-ever text-guided model general tasks. With carefully designed ablation experiments, investigate influence various augmentation ratios, hyper-parameter settings, patch size generating merging mask combined different network architectures.