3D CSFA-UNet: a unified attention-driven deep learning framework for accurate knee MRI segmentation and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopaedics

Background:

• Current multi-task deep learning methods for joint segmentation and classification of 3D knee MRIs have limitations in clinical applicability.
• These limitations include restricted multi-scale context modeling, insufficient attention to spatial-channel cues, and high computational costs.

Purpose of the Study:

• To propose a unified, multi-stage framework for joint segmentation and classification of 3D knee MRI volumes.
• The goal is to improve diagnostic precision, interpretability, and efficiency for clinical applications.

Main Methods:

• The framework utilizes Gaussian Guided Filtering for noise suppression and boundary enhancement.
• A 3D CSFA-UNet (Channel-Spatial Feature Attention) with Atrous Spatial Pyramid Pooling (ASPP) is employed for segmentation.
• A Spiking Transformer network with Leaky Integrate-and-Fire (LIF) neurons and graph-attention layers is used for classification, preceded by a Desert Scorpion Feature Selector (DSFS) for feature selection.
• Hyperparameter tuning is performed using Falcon Hunting Optimisation (FHO).

Main Results:

• Segmentation performance achieved a Dice Similarity Coefficient (DSC) of 98.10%, Intersection over Union (IoU) of 96.26%, Average Surface Distance (ASD) of 0.45 mm, and 95th percentile Hausdorff Distance (Hd95) of 1.85 mm.
• Classification accuracy reached 99.15%, with precision of 98.82%, recall of 99.11%, and an F1-score of 99.04% on the OAI dataset.
• The model demonstrated robustness and reliability in both segmentation and grading tasks.

Conclusions:

• The proposed framework offers a clinically relevant, interpretable solution for joint segmentation-classification of 3D knee MRIs.
• This approach advances image-guided orthopaedic diagnostics by improving precision and efficiency.
• The study highlights the potential of this unified framework for enhanced diagnostic capabilities in clinical settings.