Differentiation of malignant from benign pleural effusions based on artificial intelligence

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pleural effusion diagnosis can be challenging.
• Distinguishing benign pleural effusion (BPE) from malignant pleural effusion (MPE) is critical for patient management.
• Current diagnostic methods may be invasive or time-consuming.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) models for segmenting and classifying pleural effusions using thoracic CT images.
• To differentiate between benign pleural effusion (BPE) and malignant pleural effusion (MPE) with high accuracy.
• To improve the efficiency and accuracy of pleural effusion diagnosis.

Main Methods:

• Constructed a segmentation model (M1) by combining 3D spatially weighted U-Net with 2D classical U-Net.
• Developed a classification model (M2) using CT volumes and 3D pleural effusion masks.
• Utilized a training cohort (607 cases), an internal testing cohort (311 cases), and an external testing cohort (362 cases).

Main Results:

• The segmentation model (M1) achieved superior performance metrics compared to baseline U-Net models.
• The classification model (M2) demonstrated an area under the receiver operating characteristic curve of 0.842 in the external testing cohort.
• Inputting CT volume with 3D pleural effusion masks significantly improved classification performance for differentiating BPE and MPE.

Conclusions:

• Deep learning models can effectively segment pleural effusions on CT scans.
• The developed AI model shows encouraging performance in the differential diagnosis of benign pleural effusion (BPE) and malignant pleural effusion (MPE).
• This AI approach holds potential for non-invasive and accurate pleural effusion characterization.