Canine Pneumothorax Diagnosis using LLM-Guided Segmentation Refinement and Classification

ICLR 2026 Conference Submission16833 Authors

Keywords: image segmentation, LLM

Abstract: Automated diagnosis of canine pneumothorax is a critical veterinary emergency, challenged by the scarcity of large-scale annotated datasets and the need for high-precision, trustworthy models. To address these challenges, we first introduce a novel, publicly available dataset for canine pneumothorax with expert-provided, pixel-level annotations. Based on this dataset, we propose an innovative diagnostic framework guided by a Large Language Model (LLM). The framework begins with a data-efficient segmentation model (UnetFlowMatch) generating an initial mask. Subsequently, we pioneer a closed-loop iterative refinement process where the LLM assesses the original image with the mask overlaid and generates specific, natural language instructions. These instructions guide the segmentation model through corrective iterations until high precision is achieved. Finally, a definitive classification of pneumothorax presence is performed on the focused lesion area, which is delineated by the final, refined mask. This novel paradigm, which integrates an LLM as a "supervisor" deep within the segmentation loop, significantly improves segmentation accuracy. It also enhances the trustworthiness of the final diagnosis by ensuring it is based on a high-precision, validated region. Our work not only contributes a valuable dataset but also pioneers a new path for leveraging LLMs to improve the accuracy and reliability of AI diagnostic systems.

Primary Area: datasets and benchmarks

Submission Number: 16833