This repository will host the implementation of the DTSA framework, a novel approach designed to enhance semi-supervised medical image segmentation through dual-textual guidance.
Semi-supervised Medical Image Segmentation (SMIS) often faces challenges such as limited annotated data and model cognitive biases. DTSA addresses these issues by leveraging two complementary types of textual cues to improve the model's contextual understanding:
Class-aware Cues: Utilizes fixed templates (e.g., "A pixel of [class]") to provide explicit category-specific attributes.
Global-context Cues: Incorporates instance-level descriptions (e.g., size, shape, and appearance) to capture comprehensive spatial features.
The DTSA framework integrates vision and language through three key components:
Class-aware Cognitive Enhancement (CCE): Combines visual features with class-specific text cues to enhance feature representation.
Global-context Cognitive Enhancement (GCE): Augments visual embeddings with global textual cues, using a retrieval-based approach for unlabeled samples.
Interactive Cognitive Calibration (ICC): Facilitates mutual learning between the two types of text-enhanced features to mitigate cognitive biases.
The full source code, pre-trained models, and configuration files will be released upon the official acceptance of the paper. Stay tuned for updates!