Multi-to-binary network (MTBNet) for automated multi-organ segmentation on multi-sequence abdominal MRI images.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Fully convolutional neural networks (FCNs) excel at semantic segmentation but struggle with multi-object segmentation, particularly in medical imaging where organ scales vary significantly.
• Existing segmentation networks often treat all organs equally, which is suboptimal for tasks involving diverse organ sizes and appearances.

Purpose of the Study:

• To address the limitations of FCNs in multi-organ segmentation by proposing a novel network architecture.
• To improve the segmentation accuracy of small or challenging organs within complex medical images.

Main Methods:

• Developed a multi-to-binary network (MTBNet) that decomposes multi-organ segmentation into multiple binary segmentation tasks.
• Introduced a plug-and-play multi-to-binary (MTB) block featuring parallel branches with varying convolutional layers and a probability gate (ProbGate).
• The ProbGate predicts class existence, supervised by an auxiliary loss, enabling adaptive feature weighting for improved prediction.

Main Results:

• The MTBNet demonstrated improved segmentation accuracy for small organs on a challenging abdominal MRI dataset.
• The proposed method effectively adjusted feature influence, reducing the incidence of missed or misidentified organs compared to traditional methods.
• The MTB block's ability to adapt feature maps enhanced overall segmentation performance.

Conclusions:

• The MTBNet offers a more effective approach to multi-organ segmentation by converting the task into binary problems.
• The adaptive feature adjustment mechanism, particularly the ProbGate, is crucial for handling variations in organ scale and appearance.
• This method shows significant potential for improving diagnostic accuracy in medical image analysis.