Lesion-aware cross-phase attention network for renal tumor subtype classification on multi-phase CT scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multi-phase computed tomography (CT) is crucial for non-invasive kidney cancer diagnosis.
• Radiologist assessment of renal lesion enhancement patterns across CT phases shows inter-observer variability.
• Existing deep learning models for kidney cancer diagnosis do not explicitly leverage CT phase relationships.

Purpose of the Study:

• To develop a novel deep learning network, LACPANet, for accurate classification of kidney cancer subtypes.
• To effectively capture temporal dependencies of renal lesions across multiple CT phases.
• To address the limitations of current diagnostic methods by modeling inter-phase enhancement patterns.

Main Methods:

• Proposed a lesion-aware cross-phase attention network (LACPANet) for time-series multi-phase CT images.
• Introduced a 3D inter-phase lesion-aware attention mechanism to learn lesion features and inter-phase relations.
• Implemented a multi-scale attention scheme to aggregate temporal patterns at different spatial scales.

Main Results:

• LACPANet effectively captures temporal dependencies of renal lesions across CT phases.
• The 3D attention mechanism learns discriminative inter-phase enhancement patterns.
• The multi-scale attention scheme further enhances the aggregation of temporal lesion features.
• Extensive experiments demonstrated superior diagnostic accuracy compared to state-of-the-art approaches.

Conclusions:

• LACPANet offers a significant advancement in the differential diagnosis of kidney cancer using multi-phase CT.
• The proposed attention mechanisms effectively model inter-phase relationships for improved diagnostic performance.
• This approach has the potential to reduce inter-observer variability and enhance clinical decision-making in kidney cancer diagnosis.