# A deep learning model based on ultrasound imaging to differentiate malignant from benign pleural effusion: a multicenter cohort study

## TL;DR

A deep learning model using ultrasound images can help distinguish between malignant and benign pleural effusion, potentially reducing the need for invasive procedures.

## Contribution

The study introduces a deep learning model for non-invasive diagnosis of malignant pleural effusion using ultrasound imaging.

## Key findings

- The model achieved an accuracy of 0.750 in internal testing and 0.774 after fine-tuning with external data.
- The model's sensitivity and specificity were 0.710 and 0.803 in internal testing, and 0.818 and 0.611 in external testing.
- The model shows potential as a non-invasive diagnostic tool for pleural effusion.

## Abstract

Thoracentesis is required for malignant pleural effusion (MPE) diagnosis. However, it is an invasive procedure and carries risks. It remains unknown whether deep learning using ultrasound images could become a non-invasive diagnostic approach.

Patients with pleural effusion detected by thoracic ultrasound and received diagnostic thoracentesis were retrospectively collected from two sites. The internal cohort collected patients from the National Taiwan University Hospital (NTUH) Hsin-Chu branch (2014–2021), whereas the external cohort collected patients from NTUH (2020–2021). The MPE was confirmed by cytopathology reports, while benign pleural effusion was ascertained by negative cytology and compatible clinical courses. A convolutional deep learning model was used to identify MPE. Performance metrics included accuracy, F1 score, sensitivity, specificity and the area under the receiver operating characteristic curve (AUC).

A total of 601 and 144 patients from the internal cohort and the external cohort were used for model development. The model achieved promising results in internal testing (accuracy = 0.750 [95% CI: 0.689–0.811], sensitivity = 0.710 [95% CI: 0.619–0.798], specificity = 0.803 [95% CI: 0.704–0.893], F1 = 0.763 [95% CI: 0.691–0.826], AUC = 0.814 [95% CI: 0.746─0.873]). After fine-tuning with small number of external images, the model achieved the following performance on the external testing set: accuracy = 0.774 [95% CI: 0.679–0.857], sensitivity = 0.818 [95% CI: 0.723–0.905)], specificity = 0.611 [95% CI: 0.389–0.846], F1 = 0.850 [95% CI: 0.776–0.913], AUC = 0.753 [95% CI: 0.596–0.885].

Our deep learning model holds promise as a non-invasive point-of-care modality for assistance in pleural effusion diagnosis.