Case study: differentiating interstitial lung disease subtypes using AI-based image analysis

Interstitial lung disease is a large group of diseases that cause scarring or fibrosis of the lungs, stiffening them, making breathing difficult, and affecting oxygen distribution in the bloodstream.¹ These diseases are complex and have a lot of histologic overlap, requiring pathologists to develop tools that can accurately differentiate between the various subtypes to improve pathology reporting, diagnosis, and clinical decision-making. 

Dr. Maxwell L. Smith, a pathologist at Mayo Clinic, developed AI models that recognize and quantitate unique histologic features of interstitial lung disease in an attempt to differentiate between ILD subtypes. This case study summarizes Dr. Smith’s presentation at Aiforia’s webinar, “Using AI in anatomic pathology: application to interstitial lung disease and liver transplantation.”   

Three major fibrotic interstitial lung diseases may present with geographically and temporally heterogeneous fibrosis 

Interstitial lung disease subtypes like connective tissue disease (CTD), chronic hypersensitivity pneumonitis (CrHP), and idiopathic pulmonary fibrosis (IPF) may present as usual interstitial pneumonia (UIP) and have significant histologic overlap. 

The geographic heterogeneity of UIP in these ILD biopsies is characterized by “patchy” regions of pink, scarred tissue that alternate with the thin delicate alveolar walls of normal lung parenchyma. However, UIP also shows temporal heterogeneity in that there are areas of active fibroplasia (fibroblastic foci) in addition to areas of advanced scarring. 

“The similarity of UIP patterns across CTD, CrHP, and IPF makes it challenging to differentiate between the ILDs and accurately diagnose a patient with one but not the others without relying on additional histologic features,” Dr. Smith explained.

The need for lung biopsies to diagnose ILD subtypes

The current histologic criteria for the diagnosis of fibrotic ILD is complex, focusing on diagnosing IPF but not necessarily CTD or CrHP associated UIP. Despite multidisciplinary diagnosis (MDD) being the gold standard, the field has shifted away from lung biopsies,  relying heavily on radiographic features, mainly because traditional surgical pathology does not add sufficient value to justify the risk.

“Historically, we are not adding enough value to justify the risk of lung biopsy. However, there are important histologic features in these biopsies that deserve closer attention than the field has given them in the past.” – Maxwell L. Smith, MD, Pathologist, Mayo Clinic

For instance, a 2012 study showed statistically significant differences between histologic features that distinguish CrHP-UIP from IPF-UIP.²  

It’s critical to differentiate between ILD subtypes

Because of prognostic and treatment differences, Dr. Smith emphasized the need for pathologists to establish clear guidelines for distinguishing between ILD subtypes. For instance, the five-year mortality rate for IPF is much higher than that of CTD and CrHP.³ 

Likewise, an IPF patient who receives treatment developed for CTD is at a higher risk for death or hospitalization, underscoring the need for a multidisciplinary diagnosis to differentiate between the ILD subtypes accurately.⁴

AI can help accurately and reproducibly quantitate lung biopsy histologic features

However, challenges like large lung tissue biopsies make it difficult for human pathologists to accurately and reproducibly evaluate the extent of a histologic finding. That’s where AI capabilities step in to offer potential insights into prognostic differences.

Dr. Smith used Aiforia® Create, an AI development tool, to build two unique AI models that recognize most of the important histologic features required for pathologic ILD diagnosis. Upon training and validation, these models performed well and could accurately recognize quantitate regions of advanced fibrosis and fibroblastic foci compared to those without fibrosis in the lung biopsies. 

The AI models could also identify histologic features like fibrin, granulomas, bone fragments, inflammation, lymphoid follicles, and germinal centers, which are critical to diagnosing ILD subtypes, including differentiating connective tissue disease and chronic hypersensitivity pneumonitis from idiopathic pulmonary fibrosis.

The AI model recognizes areas containing fibrosis (highlighted in green mask)

AI models can support pathology reporting

Dr. Smith’s group plans to apply the AI model they have developed to pristine, gold-standard MDD diagnostic cases to confirm the expected differences they observe based on those diagnoses. The team will also evaluate associations between these differences and their associated prognostic or therapeutic responses.

Dr. Smith believes incorporating AI-driven prognostic-, diagnostic-, or therapeutic-related scores will provide the decision support needed to deliver effective patient care. For instance, FF activity scores may reveal a patient’s risk of disease progression, whereas inflammatory density scores may indicate a patient’s likelihood to respond to immunosuppression. 

“We feel good about the AI model we have developed and validated to recognize and quantitate histologic features of interstitial lung disease. There’s a gold mine of quantitative data buried in tissue specimens, and we are just now beginning to mine for it,” Dr. Smith added.

Aiforia® Create is the most versatile tool for developing, customizing, and validating deep learning AI models for histological features and patterns in image analysis.

Learn more about the interstitial lung disease AI models: Watch the full webinar recording

References