CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning

Name: CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning
Uploaded: 2026-10-13T14:20:48.757369+00:00
Duration: 293 s
Description: CORE-Seg is a new framework that combines Multimodal Large Language Models (MLLMs) with reinforcement learning to segment complex medical lesions. It uses a two-stage training process, starting with supervised fine-tuning and moving to reinforcement learning guided by a dual-granularity reward mechanism. This approach achieves state-of-the-art segmentation accuracy on challenging lesions.

CORE-Seg is a new framework that combines Multimodal Large Language Models (MLLMs) with reinforcement learning to segment complex medical lesions. It uses a two-stage training process, starting with supervised fine-tuning and moving to reinforcement learning guided by a dual-granularity reward mechanism. This approach achieves state-of-the-art segmentation accuracy on challenging lesions.

Open the full episode