CASE STUDY
Early Diagnosis and Prediction of Pulmonary Fibrosis
Abstract
The aim of this work is to identify an automatic, accurate, and fast deep learning segmentation approach, using a very small dataset of high-resolution computed tomography images of patients with idiopathic pulmonary fibrosis. In this way, we aim to enhance the methodology performed by healthcare operators in radiomics studies where operator-independent segmentation methods must be used to correctly identify the target and, consequently, the texture-based prediction model. Conclusions: We demonstrated that deep learning models can be efficiently applied to rapidly segment and quantify the parenchyma of patients with pulmonary fibrosis, in order to produce user-independent results.
Introduction
Despite the recent update of the histological criteria of idiopathic pulmonary fibrosis (IPF), few prognostic histological factors have been identified thus far. Fibroblast foci (FF) are key histological features in IPF, which manifests histologically as the usual interstitial pneumonia (UIP) pattern. High numbers of FF have been associated with worse outcomes for patients with IPF in several studies, but some controversial results have also been published.
The development of artificial intelligence (AI) enables new approaches to image analysis. AI models have been shown to recognize histological UIP pattern by using genomic data from lung biopsies. Radiological findings can also be quantitated using automated image analysis and have been associated with pulmonary function, survival, and response to antifibrotic medication. In a manner comparable to radiologists, an AI model can classify fibrotic lung diseases according to high-resolution computed tomography images. AI models have been used in the histology of experimental mouse models of pulmonary fibrosis. To our knowledge, histological features of IPF samples have not been previously studied using automated image analysis. Before developing diagnostic AI models for the UIP pattern, the ability of AI to identify specific histological features should be tested.
We aim to test the previous association between FF and prognosis of patients with IPF using the automated image analysis. Our approach was to pilot an AI model with a small data set and test its generalizability in slides that were not included in the training data set. Using lung tissue samples of thoroughly characterized patients from the FinnishIPF registry patients, we develope the AI model with a deep convolutional neural network (CNN).
Elementary Data Analysis of patients
