Automated Interpretation of Lung Ultrasound for COVID-19 and Tuberculosis diagnosis

Abstract

BACKGROUND. Early and accurate detection of infectious respiratory diseases like COVID-19 and tuberculosis (TB) plays a crucial role in effective management and the reduction of preventable mortality. However, molecular diagnostic tests for these infections are expensive and not easily implementable in resource-limited settings, which suffer the majority of the burden. Lung Ultrasound (LUS) presents a cost-effective alternative for disease detection at the point of care, and its potential can be enhanced through automation using deep learning techniques to overcome the challenges of difficult image interpretation. DeepChest, a neural attention network, has been designed to predict the diagnosis of COVID-19 from LUS images and has shown promising results.

AIM. This study aims to further explore the predictive capabilities of DeepChest with an out-of-distribution dataset and extend its application to TB diagnosis.

METHODS/FINDINGS. For COVID-19 (resp. TB), this study is based on a main dataset and an out-of-distribution dataset consisting of patients attending an emergency department in Switzerland (resp. an outpatient facility in a TB-endemic region) between February 2020 and March 2021 (resp. between October 2021 and May 2023) with suspected COVID-19 (resp. TB) pneumonia and ground truth labels are RT-PCR.
To assess the generalizability of DeepChest for COVID-19 diagnosis, the model trained on the main dataset (296 patients, still LUS images) was tested on an out-of-distribution dataset (135 patients more severely affected, mainly frames extracted from LUS videos). We found that the performance on the out-of-distribution dataset was poor. However, by fine-tuning DeepChest on the latter, the best performance was achieved when using three random frames per video (AUC ROC 0.84 +/- 0.03) instead of one (AUC ROC 0.78 +/- 0.06).
To assess the performance of DeepChest for TB diagnosis, the model was trained on the main dataset (386 patients, still LUS images collected in an urban area of Benin). Its performance (AUC ROC 0.92 +/- 0.02) outperformed the LUS expert baseline (AUC ROC 0.84 +/- 0.01) and the clinical baseline (AUC ROC 0.89 +/- 0.01). Additionally, a multimodal model incorporating clinical data alongside LUS images was developed. It achieved the best classification performance (AUC ROC 0.94 +/- 0.01). However, when tested on an out-of-distribution dataset (150 patients, still LUS images collected in a rural region of South Africa, with much more severe presentation), the generalizability of DeepChest was found to be low (AUC ROC 0.64 +/- 0.03).

CONCLUSION. The findings of this study are promising, demonstrating the potential of DeepChest for COVID-19 and TB diagnosis using LUS images. Poor generalization to populations with more severe forms of the diseases shows the importance of either collecting more representative samples or ensuring that implementation is constrained to the target population.

Popular Abstract

Lung ultrasound (LUS) could provide accessible and accurate diagnosis for lung diseases in resource-limited environments, leveraging deep learning to improve difficult image interpretation.

Despite being a treatable disease, tuberculosis (TB) is still wreaking havoc around the world, highlighting the importance of early and swift diagnosis to reduce mortality. This project has demonstrated the effectiveness of a deep learning model in accurately predicting TB diagnosis from LUS images.
The work is based on a pre-existing deep learning model - DeepChest - originally designed to predict the diagnosis of COVID-19 from LUS images. Unlike many studies on this subject, DeepChest makes predictions at the level of the patient and not of the image itself. That is, DeepChest exploits the set of available images belonging to a patient to predict his or her diagnosis, i.e., whether he or she is positive or negative for the disease. A previous study demonstrated the effectiveness of DeepChest on a dataset consisting of still LUS images (selected by an LUS expert) belonging to almost 400 patients with suspected COVID-19 infection.

The first part of the study was devoted to exploiting another dataset collected from patients suspected of having COVID-19. However, this new dataset contains LUS videos and not still images selected by an expert. As DeepChest only takes a selection of images input as, we had to think about a strategy for selecting the frames in the videos, bearing in mind that for reasons of computational resources, we could not use all the frames of a video as input. The naive strategy of using one random frame per video gave relatively poor results. We were able to improve slightly the results using three random frames per video but not when we tried to build synchronous sequences of frames (i.e., by making sure that for all LUS videos, we would select three consecutives frames with the same intensity and direction). This preliminary study paves the way for further research into how best to exploit LUS videos when no specific medical knowledge is available to select the best frames.

The main part of the study sought to extend the work that had previously been done on DeepChest in the context of COVID-19 diagnosis, for the diagnosis of TB. To this end we had two datasets of LUS images (selected by a LUS expert) at our disposal: a main dataset collected in an urban region of Benin (386 patients with suspected TB) and an external validation dataset collected in a rural region of South Africa (150 patients with suspected TB). These datasets represented distinct populations, with the South African region having a higher prevalence of HIV. The performance of DeepChest on the main dataset was compared with two reference models, namely a clinical model and a LUS expert model. The clinical model was obtained from tabular clinical data (age, sex, weight, height, past/present medical history, symptoms, blood pressure, etc.) and the LUS expert model was obtained from severity scores assigned to each LUS image by an expert LUS clinician. We found that DeepChest performed better than these two reference models. However, when we tested the ability of DeepChest to generalize to the patients in South Africa, the performance was poor, highlighting the dependence of the model on the population on which it is trained.

The results of this study are promising and encourage further research in this area. Particular care must be taken to ensure the quality and relevance of the datasets, which must reflect real-life scenarios as closely as possible (a TB vs. healthy classification has no clinical relevance, in contrast to a TB vs. other lung disease classification). In addition, given the stakes involved, work needs to be done on the interpretability of the models (it must be possible to understand and explain their decisions) if we want to deploy them for medical use.

• author: Soormally, Chloé
• supervisor: Mikael Nilsson; Mary-Anne Hartley
• organization: Mathematics (Faculty of Engineering)
• course: FMAM05 20231
• year: 2023
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Tuberculosis, COVID-19, Lung Ultrasound, Computer-aided detection (CAD), Deep learning
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMA-3514-2023
• ISSN: 1404-6342
• other publication id: 2023:E45
• language: English
• id: 9127725
• date added to LUP: 2023-06-26 14:42:11
• date last changed: 2023-06-26 14:42:11