FARNet

FARNet automates detection of anatomical landmarks in radiographs to support clinical diagnosis, treatment planning, and research.

Key Features:

• Encoder-Decoder Structure: An encoder-decoder architecture with a backbone pre-trained on natural images transfers learned features to mitigate limited annotated medical datasets.
• Multi-Scale Feature Aggregation Module: A decoder module performs multi-scale feature fusion to capture information at different resolutions for landmarks of varying size and appearance.
• Feature Refinement Module: A refinement module produces high-resolution heatmap regression to improve precision and reliability of landmark localization.
• Coarse-to-Fine Supervisions: Coarse-to-fine supervisory signals are applied to aggregation and refinement modules to progressively enhance detail during end-to-end training.
• Exponential Weighted Center Loss Function: An exponential weighted center loss for heatmap regression emphasizes pixel losses near landmarks while down-weighting distant pixels to improve localization accuracy.

Scientific Applications:

• Cephalometric radiographs: Evaluation on cephalometric datasets for automated craniofacial landmark detection and analysis.
• Hand radiographs: Application to hand X-rays for precise detection of skeletal landmarks relevant to orthopedic assessment.
• Spine radiographs: Use with spine radiographs for vertebral and spinal landmark localization in clinical and research studies.

Methodology:

FARNet employs an encoder-decoder architecture with a backbone pre-trained on natural images, multi-scale feature fusion in the decoder, a feature refinement module for high-resolution heatmap regression, coarse-to-fine supervisions, and an Exponential Weighted Center Loss for heatmap regression.