A Multimodal Vision-Language Framework for the Diagnosis of Farmer's Lung (Chronic Hypersensitivity Pneumonitis)

Overview

Chronic Hypersensitivity Pneumonitis (CHP) is a progressive interstitial lung disease caused by repeated inhalation of environmental antigens and is particularly prevalent among agricultural workers. Differentiating CHP from other interstitial lung diseases (ILDs) remain challenging due to overlapping radiographic and clinical features, often leading to delayed diagnosis and progression to irreversible pulmonary fibrosis. Multidisciplinary team (MDT) evaluation is the current diagnostic gold standard; however, it is time-consuming and resource intensive. This study presents AgriLungAI, a multimodal vision–language model (VLM) designed to improve the early and accurate diagnosis of CHP by integrating high-resolution computed tomography (HRCT) imaging with structured clinical exposure data. A retrospective dataset of 100 patients with confirmed ILD diagnoses was assembled, with ground-truth labels derived from radiologist interpretations and final MDT consensus. Electronic medical records were reviewed to extract exposure history and symptom data. HRCT chest scans were processed as volumetric video inputs. The model was built on Qwen2.5-VL and implemented in Python using PyTorch, with both vision and language encoders fine-tuned via Low-Rank Adaptation (LoRA). The AgriLungAI model achieved over 80% predictive accuracy and demonstrated an enhanced ability to integrate radiologic patterns with clinical context, successfully distinguishing CHP from other ILDs. These findings support the hypothesis that multimodal AI can improve diagnostic reliability and enable earlier detection of CHP, particularly in high-risk agricultural regions with limited specialist access. AgriLungAI shows promise as both a scalable clinical decision-support tool and an educational resource for training medical students and residents.