CCS-Net

CCS-Net detects hypopharyngeal cancer (HPC) and associated risk areas, including swollen lymph nodes, in T2-weighted magnetic resonance imaging (MRI) using a cascade detection deep-learning network to improve lesion localization and detection accuracy.

Key Features:

• Cascade Detection Network: Core architecture that integrates specialized components to optimize detection of small and irregular HPC risk areas in MRI.
• Convolution Kernel Switch (CKS) Block: Enables adaptive switching between standard convolution and deformable convolution within network layers to handle irregular object shapes without substantially increasing computation.
• Statistics Optimal Anchors (SOA) Block: Automatically generates optimal anchors from the statistical distribution of object sizes in the dataset to improve localization and detection accuracy.

Scientific Applications:

• HPC MRI analysis: Detection and localization of hypopharyngeal cancer lesions and swollen lymph nodes in T2-weighted MRI slices for computer-aided diagnosis studies.

Methodology:

Trained on over 1800 T2-weighted MRI slices using a cascade detection network with a CKS block (switching between standard and deformable convolution) and an SOA block for anchor generation, reporting an average precision (AP_50) of 78.90% compared with existing methods.