Multimodal AI system improves accuracy in identifying interstitial pneumonia

The multimodal artificial intelligence (AI) algorithm that combines computed tomography (CT) and histopathological images improves diagnostic accuracy, consistency, and confidence of pathologists in identifying usual interstitial pneumonia (UIP) in patients with resected lung specimens, reports a study.

“The extensive validation process, including comparisons with expert pathologists and general practitioners, highlights the significant potential of this AI tool in enhancing diagnostic capabilities and confidence among pathologists, even in cases where specialized expertise is limited,” the investigators said.

In this study, the investigators obtained a dataset of CT and pathological images from 324 patients with interstitial lung disease (ILD) between 2009 and 2021. They trained the CT component of the model to recognize 28 different radiological features. The pathological counterpart was previously developed in an earlier study.

Overall, 114 samples were selected and used in testing the multimodal AI model. Comparisons with expert and general pathologists were carried out to assess the performance of the multimodal AI.

The multimodal AI significantly enhanced the pathologists’ accuracy in distinguishing UIP from non-UIP, with an area under the curve (AUC) of 0.92. Among general pathologists, the diagnostic agreement rate significantly improved to a postmodel κ score of 0.737 from 0.273 premodel integration. [Respirology 2025;30:726-735]

Furthermore, the diagnostic consensus rate with expert pulmonary pathologists rose from κ scores of 0.278‒0.53 to 0.474‒0.602 postmodel integration. The AI model also improved diagnostic confidence among general pathologists.

“We believe that the model's performance in improving diagnostic accuracy and consistency signifies a critical step toward more standardized and reliable diagnoses in the field of ILD,” the investigators said. “Beyond the context of UIP diagnosis, this multimodal approach holds promise for addressing the complexities of other intricate medical conditions.”

MIXTURE model

In a previous study, the investigators developed the MIXTURE model, which can identify histological UIP. [Modern Pathology 2022;35:1083-1091]

“In our ongoing efforts to enhance the accuracy of the MIXTURE model, we have now introduced a novel approach: a multimodal AI that incorporates radiographic image characteristics to aid in diagnosing histological UIP.”

Combining clinical observations, radiographic data, and genomic characteristics with histopathological images results in improved diagnostic accuracy. [BMC Medicine 2012;10:100; Radiology 2017;285:12-15]

Recent advancements in machine learning have led to the creation of AI models that integrate diverse diagnostic inputs to improve precision. One example is the combination of pathological and radiographic data to improve performance in glioma diagnosis and more accurate prognostication in high-grade serous carcinoma of the ovary. [Nature Cancer 2022;3:723-733; Scientific Reports 2022;12:6111]

“In ILDs, a multidisciplinary approach is essential, with radiographic images playing a key role in providing a comprehensive view of lesions,” the investigators said.

“Integrating a chest CT AI model into the MIXTURE model, which focuses on histopathological morphology, enables a holistic approach that improves the accuracy of histological UIP diagnosis and holds promise for other complex medical conditions,” they added.