Automated interpretation and analysis of bronchoalveolar lavage fluid
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.
Section snippets
Sample processing and dataset preparation
In this study, bronchoalveolar lavage (BAL) was performed, and BALF was sampled in accordance with the standard procedure recommended in the American Thoracic Society clinical practice guideline [5]. There were no exclusion criteria, such as gender, age, or race. For each subject, 2 mL were sampled and centrifuged at 300g for 10 min, followed by resuspension with 0.5 mL normal saline. To standardize the variables in sample processing, 20 μl BALF were smeared evenly within a fixed spot on each
Results
A workstation computer with dual Intel Xeon Gold 5122 processors, 64 GB RAM, and an NVIDIA Quadro P5000 computing card was used for algorithm training. The batch size was 4, and the training time was approximately 2 days and 3 h. Fig. 4 shows some examples of object classification, where the objects detected were framed by red rectangle surroundings with predictions and chances on the upper right side. The performance of the proposed model was evaluated in the testing set, consisting of 462
Discussion
Here, we propose an approach to reveal the nature and state of lung disease rapidly, on-site, and precisely in critically ill patients in the ICU. The great performance in the testing set showed its effectiveness in automatically interpreting BALF cytology, and our model outperformed experienced practitioners when tested at the real-world patient level.
Similar studies have been performed to detect blood cells using CNN-based algorithms. Kutlu et al. tested several RCNN models’ performance in
Conclusions
In this study, we propose a model that can automatically provide the cytological background of the BALF. The excellent performance in the test set shows that our model has the potential to augment clinical decision-making for clinicians. Our model has been used in the RICU, and we are exploring more possibilities to benefit patients.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This study was funded by the Key Projects of Military Logistics Scientific Research Program (BLB18J008).
References (31)
- et al.
Assessment of the worldwide burden of critical illness: the intensive care over nations (ICON) audit
Lancet Respir. Med.
(2014) - et al.
Identifying medical diagnoses and treatable diseases by image-based deep learning
Cell
(2018) - et al.
Nasal cytology with deep learning techniques
Int. J. Med. Inform.
(2019) - et al.
Image analysis and machine learning for detecting malaria
Transl. Res.
(2018) - et al.
White blood cells detection and classification based on regional convolutional neural networks
Med. Hypotheses
(2020) - et al.
Community-acquired pneumonia
Lancet
(2015) - et al.
Rapid on-site evaluation of transbronchial aspirates in the diagnosis of hilar and mediastinal adenopathy: a randomized trial
Chest
(2011) - et al.
International study of the prevalence and outcomes of infection in intensive care units
JAMA
(2009) - et al.
The impact of hospital and ICU organizational factors on outcome in critically ill patients: results from the Extended Prevalence of Infection in Intensive Care study
Crit. Care Med.
(2015) - et al.
High levels of inherent variability in microbiological assessment of bronchoalveolar lavage samples from children with persistent bacterial bronchitis and healthy controls
Pediatr. Pulmonol.
(2020)
An official American Thoracic Society clinical practice guideline: the clinical utility of bronchoalveolar lavage cellular analysis in interstitial lung disease
Am. J. Respir. Crit. Care Med.
British Thoracic Society guideline for diagnostic flexible bronchoscopy in adults: accredited by NICE
Thorax
Automated interpretation of blood culture gram stains by use of a deep convolutional neural network
J. Clin. Microbiol.
Dermatologist-level classification of skin cancer with deep neural networks
Nature
Deep neural networks offer morphologic classification and diagnosis of bacterial vaginosis
J. Clin. Microbiol.
Cited by (5)
A comparative study of attention mechanism based deep learning methods for bladder tumor segmentation
2023, International Journal of Medical InformaticsAn improved Yolov5s based on transformer backbone network for detection and classification of bronchoalveolar lavage cells
2023, Computational and Structural Biotechnology JournalRapid on‑site pathogen diagnosis of intensive care unit: Microbiological Rapid On‑Site Evaluation
2023, Chinese Journal of Laboratory MedicineClinical update in critical care of pulmonary medicine 2022
2023, Zhonghua Jiehe he Huxi Zazhi / Chinese Journal of Tuberculosis and Respiratory Diseases