MultiResUNet : Rethinking the U-Net architecture for multimodal biomedical image segmentation.

Area of Science:

• Deep Learning in Medical Imaging
• Computer Vision
• Artificial Intelligence in Healthcare

Background:

• Deep Learning (DL) has revolutionized Medical Image Segmentation.
• U-Net is a widely adopted DL architecture in medical imaging.
• Classical U-Net shows limitations in segmenting challenging multimodal medical images.

Purpose of the Study:

• To propose modifications to the U-Net architecture for improved performance.
• To introduce MultiResUNet as a potential successor to U-Net.
• To evaluate the efficacy of MultiResUNet on diverse medical imaging datasets.

Main Methods:

• Development of a novel architecture, MultiResUNet, based on U-Net modifications.
• Extensive experimentation and comparison of MultiResUNet against the classical U-Net.
• Evaluation on five distinct multimodal medical image datasets with varying challenges.

Main Results:

• MultiResUNet demonstrates remarkable performance gains on challenging medical images.
• Relative performance improvements of 10.15%, 5.07%, 2.63%, 1.41%, and 0.62% were achieved across five datasets.
• Qualitative analysis reveals superior aspects of MultiResUNet not captured by quantitative metrics.

Conclusions:

• MultiResUNet offers significant improvements over classical U-Net for medical image segmentation, particularly in complex cases.
• The proposed modifications enhance the U-Net architecture's capability in handling challenging medical imaging scenarios.
• MultiResUNet presents a promising advancement for the field of medical image analysis.