Leveraging Memory for Improved Medical Image Segmentation with Limited Parameters.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep Learning (DL) offers fast, automatic segmentation, reducing clinical workload.
• Current DL segmentation models often use 2D approaches to minimize hardware needs, despite 3D data.
• This limits the full exploitation of volumetric information in medical imaging.

Purpose of the Study:

• To propose a novel DL architecture, the 2D Long Short Term Memory U-Net (2D LSTM U-Net).
• To effectively segment medical images by combining 2D U-Net segmentation strengths with Long Short Term Memory (LSTM) volumetric understanding.
• To maintain low hardware requirements while processing 3D image data.

Main Methods:

• Developed a hybrid 2D U-Net architecture incorporating Long Short Term Memory (LSTM) layers.
• The LSTM layers enable sequential data processing for volumetric understanding.
• Evaluated the 2D LSTM U-Net on CT-ORG and BraTS 2020 datasets for segmentation tasks.

Main Results:

• The 2D LSTM U-Net demonstrated effectiveness in both binary and multi-class segmentation.
• Achieved performance comparable to state-of-the-art 2D and 3D models.
• Required significantly fewer parameters: 1.6x fewer than some 2D models and 16.5x fewer than some 3D models.

Conclusions:

• The proposed 2D LSTM U-Net is an effective and efficient architecture for medical image segmentation.
• It successfully integrates 3D volumetric understanding into a computationally efficient 2D framework.
• This approach offers a promising solution for reducing hardware demands in clinical DL applications.