Deep-active-learning approach towards accurate right ventricular segmentation using a two-level uncertainty...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• The Right Ventricle (RV) is a critical prognostic factor in various diseases.
• Cardiac Magnetic Resonance Imaging (CMRI) is used for RV assessment via short-axis slices.
• Manual RV segmentation by radiologists is time-consuming and labor-intensive.

Purpose of the Study:

• To develop an automated method for RV segmentation in CMRI using Deep-Learning (DL).
• To address the need for large annotated datasets in DL by proposing a Deep-Active-Learning (DAL) approach.
• To improve the efficiency and accuracy of RV segmentation in clinical practice.

Main Methods:

• A Deep-Active-Learning (DAL) pipeline was designed for RV segmentation.
• Initial learning involved a personalized, augmented labeled dataset.
• A modified U-Net architecture was employed for efficient initial accuracy.
• A two-level uncertainty estimation technique selected complementary unlabeled data.
• Customized postprocessing incorporated epistemic uncertainty and Dense Conditional Random Fields.

Main Results:

• The DAL approach was tested on 791 public and 1230 custom CMRI images.
• The Dice coefficient improved from 0.86 to 0.91.
• The Hausdorff distance decreased from 7.55 to 7.45.
• The method demonstrated enhanced accuracy and efficiency in RV segmentation.

Conclusions:

• The proposed Deep-Active-Learning (DAL) approach effectively automates RV segmentation in CMRI.
• This method significantly improves segmentation accuracy and reduces manual workload.
• DAL offers a promising solution for leveraging large medical datasets in DL applications.