Related Experiment Video
Updated: Jun 18, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Efficient multi-stage feedback attention for diverse lesion in cancer image segmentation
Dewa Made Sri Arsa1, Talha Ilyas2, Seok-Hwan Park3
1Division of Electronics and Information Engineering, Jeonbuk National University, Republic of Korea; Department of Information Technology, Universitas Udayana, Indonesia; Core Research Institute of Intelligent Robots, Jeonbuk National University, Republic of Korea.
This study presents a novel iterative feedback mechanism for cancer lesion detection in medical images. The method enhances accuracy by refining segmentation directly from neural network architecture, outperforming existing state-of-the-art approaches.
Area of Science:
- Medical Imaging Analysis
- Computer-Aided Diagnosis (CAD)
- Artificial Intelligence in Healthcare
Background:
- Accurate cancer lesion identification is critical for Computer-Aided Diagnosis (CAD) systems.
- Lesion segmentation is challenging due to vague boundaries, noise, and appearance heterogeneity.
- Existing iterative segmentation methods often require an initial segmentation mask, limiting their application.
Purpose of the Study:
- To introduce an innovative iterative feedback mechanism for nuanced cancer lesion detection across various medical imaging modalities.
- To eliminate the need for an initial segmentation mask in iterative segmentation processes.
- To enhance the accuracy and robustness of CAD systems for cancer lesion identification.
Main Methods:
- Developed an iterative feedback mechanism deriving refinement directly from the encoder-decoder neural network architecture.
- Utilized a multi-scale feedback attention mechanism to guide and refine predicted masks in subsequent iterations.
- Implemented a weighted feedback loss function combining global and iteration-specific considerations for improved parameter estimation.
Main Results:
- The proposed method demonstrated superior performance compared to state-of-the-art techniques across colonoscopy, ultrasonography, and dermoscopic images.
- Achieved high accuracy in both standard and challenging out-of-domain segmentation tasks.
- Validated the robustness and versatility of the approach in diverse medical imaging contexts.
Conclusions:
- The novel iterative feedback mechanism offers a significant advancement in cancer lesion detection within CAD systems.
- The method's ability to refine segmentation without an initial mask enhances its practical applicability.
- The approach shows strong potential for improving diagnostic accuracy across multiple medical imaging modalities.
