SentZero: An Enhanced Sentence-Centric Vision-Language Pretraining for Multi-Task Zero-Shot Chest X-Ray Analysis
SentZero targets the prompt-alignment problem that keeps chest X-ray VL models tied to finetuning.
The paper says radiology reports are too long and dense for simple zero-shot prompts to align cleanly. Its method restructures report content at the sentence level with LLM help, expands positive-pair diversity, and adds a loss term meant to reduce false negatives from clinically equivalent sentences. It also modulates visual embeddings based on the input sentence, so image features shift with the clinical wording. The authors report stronger zero-shot generalization across multiple chest X-ray tasks and datasets than prior multi-task zero-shot methods. HF Daily Papers' note
The paper says radiology reports are too long and dense for simple zero-shot prompts to align cleanly. Its method restructures report content at the sentence level with LLM help, expands positive-pair diversity, and adds a loss term meant to reduce false negatives from clinically equivalent sentences. It also modulates visual embeddings based on the input sentence, so image features shift with the clinical wording. The authors report stronger zero-shot generalization across multiple chest X-ray tasks and datasets than prior multi-task zero-shot methods. HF Daily Papers' note
score 4