We introduce X-PLAIN, a vision-language distillation framework for interpretable chest X-ray diagnosis. A frozen BLIP-2 teacher provides rich multimodal embeddings and textual explanations, while a compact EfficientNet-B3 student learns to replicate these representations and perform multilabel disease classification. A fine-tuned GPT-2 decoder generates natural-language explanations from the student's distilled embeddings, validating semantic fidelity. Trained on a balanced 100k-pair subset of MIMIC-CXR, the student is optimized using a multi-task objective that combines classification, visual, and textual distillation losses. Evaluations on MIMIC-CXR and zero-shot transfer to OpenI show strong diagnostic accuracy and robust semantic alignment, despite the known limitations of n -gram metrics on short radiology impressions. X-PLAIN achieves an $800 \times$ parameter reduction compared to BLIP-2 with minimal accuracy loss, offering a scalable and interpretable pathway toward clinically deployable medical AI.
Paper
Full text
X-Plain: Vision-Language Knowledge Distillation for Interpretable Chest X-Ray Diagnosis
Semantic Scholar · Medicine · 2026
Abstract
We introduce X-PLAIN, a vision-language distillation framework for interpretable chest X-ray diagnosis. A frozen BLIP-2 teacher provides rich multimodal embeddings and textual explanations, while a compact EfficientNet-B3 student learns to replicate these representations and perform multilabel disease classification. A fine-tuned GPT-2 decoder generates natural-language explanations from the student's distilled embeddings, validating semantic fidelity. Trained on a balanced 100k-pair subset of MIMIC-CXR, the student is optimized using a multi-task objective that combines classification, visual, and textual distillation losses. Evaluations on MIMIC-CXR and zero-shot transfer to OpenI show strong diagnostic accuracy and robust semantic alignment, despite the known limitations of n -gram metrics on short radiology impressions. X-PLAIN achieves an $800 \times$ parameter reduction compared to BLIP-2 with minimal accuracy loss, offering a scalable and interpretable pathway toward clinically deployable medical AI.