MedDelinea: Scalable and Efficient Medical Image Segmentation via Controllable Diffusion Transformers

Gayatri Deshmukh, Onkar Kishor Susladkar, Debesh Jha, Elif Keles, Halil Ertugrul Aktas, alpay medetalibeyoglu, Daniela P. Ladner, Amir A. Borhani, Gorkem Durak, Ulas Bagci

Keywords: Diffusion Models, Diffusion Transformer, Segmentation, Volumetric Segmentation, latent diffusion

TL;DR: Latent Diffusion Based Segmentation

Abstract: We introduce MedDelinea, a novel medical image segmentation architecture that leverages a controllable module, drawing inspiration from ControlNet, within the Diffusion Transformers (DiT) framework. By doing so, we effectively address three key challenges inherent to segmentation tasks: (1) limited availability of labeled data, (2) variability in image modalities, and (3) the need for precise boundary delineation. MedDelinea is pre-trained on a large-scale medical dataset, thereby mitigating overfitting risks and enabling efficient transfer across diverse imaging scenarios with minimal fine-tuning requirements. The modular design of MedDelinea facilitates scalable and efficient computation, while maintaining high-quality segmentation performance in both supervised and zero-shot settings. Through extensive empirical evaluations on multiple datasets, we demonstrate that MedDelinea outperforms existing state-of-the-art segmentation approaches, showcasing its potential for robust and accurate medical image analysis

Primary Subject Area: Segmentation

Secondary Subject Area: Generative Models

Paper Type: Both

Registration Requirement: Yes

Reproducibility: https://github.com/Onkarsus13/MedDelinea

Visa & Travel: Yes

Submission Number: 18