MSPA-DLA++: A Multi-Scale Phase Attention Deep Layer Aggregation for Lesion Detection in Multi-Phase CT Images

Area of Science:

• Medical Imaging
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Object detection in medical images, particularly liver lesions in CT scans, faces challenges due to significant scale variations and the need for effective feature extraction.
• Convolutional Neural Networks (CNNs) show promise but require specialized architectures to handle the complexities of medical imaging data.

Purpose of the Study:

• To develop an advanced deep learning model for accurate liver lesion detection in multi-phase CT images.
• To address the challenges of scale variation and extract subtle spatial-temporal features crucial for identifying liver lesions.

Main Methods:

• Proposed MSPA-DLA++ (Multi-phase Attention with Group-based Deep Layer Aggregation) as a backbone feature extraction network.
• Implemented an anchor-free approach for liver lesion detection.
• Utilized multi-phase CT imaging and attention mechanisms to capture temporal and spatial information.

Main Results:

• MSPA-DLA++ demonstrated effectiveness on both public (LiTS2017) and private multi-phase datasets.
• The proposed method achieved a performance improvement of approximately 3.7% over existing state-of-the-art networks.
• Successfully addressed scale variations and extracted significant hidden features for improved detection.

Conclusions:

• MSPA-DLA++ offers a robust solution for liver lesion detection in multi-phase CT scans.
• The method's ability to handle scale variations and extract complex features enhances diagnostic accuracy.
• This advancement holds potential for improving early detection and treatment planning for liver diseases.