Self-adaptive forward-forward network for anomaly detection and medical image analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Robust anomaly and out-of-distribution (OOD) detection in radiology is crucial for patient safety and requires accurate, interpretable, and efficient methods.
• Conventional back-propagation models face limitations in meeting these demands simultaneously, especially under real-world distributional shifts.
• Forward-forward learning offers a resource-efficient and biologically plausible alternative but has been hindered by scalability and generalization issues in medical applications.

Purpose of the Study:

• To introduce the Convolutional Forward-Forward Algorithm (CFFA) and SaFF-AD, a self-adaptive forward-forward network for anomaly and OOD detection in medical imaging.
• To address the scalability and generalization limitations of existing forward-forward approaches for high-dimensional medical image analysis.
• To enable stable learning under constrained computational budgets and in one-shot training regimes for anomaly detection.

Main Methods:

• Developed the Convolutional Forward-Forward Algorithm (CFFA), a parameter-efficient reformulation of forward-forward learning tailored for medical images.
• Introduced SaFF-AD, a self-adaptive forward-forward network leveraging intrinsic goodness statistics for anomaly and OOD detection.
• Implemented autonomous configuration of optimization dynamics, architectural depth, and goodness normalization within SaFF-AD.

Main Results:

• SaFF-AD achieved competitive or superior anomaly detection performance compared to back-propagation models across multiple medical imaging benchmarks.
• The proposed method required substantially fewer parameters and forward evaluations than conventional models.
• The forward-forward goodness signal enabled self-supervised anomaly and OOD detection without auxiliary networks or post-hoc methods.

Conclusions:

• Forward-forward learning, via SaFF-AD, presents a viable and practical alternative to deep learning for safety-critical medical image analysis.
• SaFF-AD is particularly suitable for settings with constrained budgets, limited data, and distributional uncertainty.
• The approach offers a unified, interpretable, and efficient framework for anomaly detection, well-suited for real-world clinical deployment.