Automated interpretation and analysis of bronchoalveolar lavage fluid

Highlights

• The combination of AI and specimen morphological recognition has a promising future.

• Bronchoalveolar lavage fluid can reflect the nature of lung diseases.

• The combination of deep learning and pathologists may outperform operators.

• Automated analysis of BALF can augment clinical decisions for lung diseases.

Abstract

Background

The cytological analysis of bronchoalveolar lavage fluid (BALF) plays an essential role in the differential diagnosis of respiratory diseases. In recent years, deep learning has demonstrated excellent performance in image processing and object recognition.

Objectives

We aim to apply deep learning to the automated interpretation and analysis of BALF.

Method

Visual images were acquired using an automated biological microscopy platform. We propose a three-step algorithm to automatically interpret BALF cytology based on a convolutional neural network (CNN). The clinical value was evaluated at the patient level.

Results

Our model successfully detected most cells in BALF specimens and achieved a sensitivity, precision, and F1 score of over 0.9 for most cell types. In two tests in the clinical context, the algorithm outperformed experienced practitioners.

Conclusion

The program can automatically provide the cytological background of BALF and augment clinical decision-making for clinicians.

Introduction

According to World Health Organization’s Global Health Estimates, respiratory diseases are the second leading cause of deaths in 2019, among which lower respiratory infections remain the deadliest communicable disease, claiming 2.6 million deaths. In the intensive care unit (ICU), 51% of patients were considered infected, and 64% of the infections were of respiratory origin. Once considered infected, the ICU mortality rate doubled [1], [2], [3]. The conditions of critically ill patients vary every minute; to improve the prognosis of patients, the most important is to make a diagnosis and exclude pulmonary infections, such as aspiration pneumonia or suppurative infections [4]. If bronchoalveolar lavage fluid (BALF) cytology is indicative of an infectious pattern, then antimicrobials could be promptly administered to fight against pulmonary infections and improve prognosis, and if not, then differential diagnosis should be considered. BALF is a precious specimen that reflects the cytological nature of a lung disease. Normally, BALF is collected from high-resolution computed tomography (HRCT)-identifying target areas, such as those of alveolar ground glass opacity, fine reticulation, or more prominent nodular profusion [5], [6]. The standard sample retention process of BALF is as follows [5]: During the standard flexible bronchoscopy, when the bronchoscope is placed within the HRCT-identifying target areas, normal saline (approximately 100–300 mL) is instilled and then retrieved through the bronchoscope, by which secretions that coat the surfaces of alveolar epithelium and bronchial are obtained and distributed. BALF accounts for the cellular and acellular components of alveolar epithelium and distal bronchial. As a result, BALF could provide cytological information of pulmonary parenchyma to support diagnosis or narrow differential diagnosis at the clinical context [5], [6]. For a normal non-smoker, the proportion of alveolar macrophages is greater than 85% in BALF; the lymphocyte proportion is between 10% and 15%; the neutrophil and eosinophil proportions are less than 3% and 1%, respectively; and the squamous epithelial cell or ciliated columnar epithelial cell proportion is less than 5%. A lymphocyte proportion greater than 50% is suggestive of hypersensitivity pneumonitis or cellular nonspecific interstitial pneumonia. A neutrophil proportion greater than or equal to 50% strongly suggests bacterial infections, such as aspiration pneumonia or suppurative infection.

The interpretation of BALF is normally performed with a biological oil lens capable of magnification of 1000x, and enumeration of at least 400 cells is required. Practically, the cytological analysis of BALF is time-consuming, tiresome, and highly operator dependent, which make it difficult to meet the standards. In China, pathologists often need to read hundreds of slides in one day, increasing workload, and the lack of experienced operators potentially affect the accuracy of results [7].

In recent years, deep learning technology has shown excellent competence in image processing and object recognition in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC), and many groups have attempted to explore the possibilities of applying deep learning to medical image interpretation. Andre et al. utilized an Inception V3 CNN architecture from the 2014 ILSVRC to complete the automated classification of skin cancer and malignant melanomas [8] based on a dataset of 129,450 clinical images. The CNN achieved a performance comparable to that of the experts. Daniel et al. later applied the same model to the classification of common treatable blinding retinal diseases based on a dataset of optical coherence tomography images [9], and the AI system also showed great performance when later used for the diagnosis of pediatric pneumonia. Other applications of convolutional networks to microscopic images include bacterial morphotype identification for bacterial vaginosis diagnosis, nasal cytology interpretations, microscopic fecal image recognition, and malaria parasite automated detection [10], [11], [12], [13], [14], [15].

Here, we developed an integrated scanning and recognition system that can perform simultaneous scanning and identification of BALF images. For image scanning, we overcame common difficulties, such as missing the appropriate focal plane or narrow scanning area, and we developed HOWSOME MICROSCAN-50, which could autofocus the Diff-Quik staining of BALF slides, scan distributed positions to generate 300 images per slide, and automatically load up to 50 slides in order in one test. A flowchart of the specimen processing, automated scanning, and interpretation is shown in Fig. 1. For a cytological interpretation, we developed a three-step program composed of cell extraction and classification. After testing in a clinical context, our program achieved high accuracy and showed great potential in assisting clinicians in making clinical diagnoses.