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

Authors

Abstract

Robust anomaly and out-of-distribution (OOD) detection in radiology demands learning methods that are accurate, interpretable, computationally efficient, and reliable under real-world distributional shifts. Existing back-propagation-trained models often struggle to meet these requirements simultaneously, while forward-forward learning, despite its conceptual appeal as a resource-efficient and biologically plausible alternative, has so far seen limited adoption in image-based and safety-critical medical applications due to scalability and generalisation limitations. We revisit back-propagation-free learning for the open-world clinical setting and discuss the Convolutional Forward-Forward Algorithm (CFFA), a parameter-efficient reformulation of the Forward-Forward Algorithm tailored to high-dimensional medical image analysis. CFFA incorporates convolutional structure and layer-wise local objectives, overcoming key scalability and generalisation limitations of existing forward-forward approaches while retaining their resource-efficient training paradigm. Building on the observation that the forward-forward objective yields intrinsic and interpretable goodness statistics that directly quantify conformity to the learned data distribution, we introduce SaFF-AD, a self-adaptive forward-forward network explicitly designed for anomaly and OOD detection. SaFF-AD autonomously configures optimisation dynamics, architectural depth, and goodness normalisation, enabling stable learning under constrained computational budgets and in one-shot training regimes. Extensive experiments across multiple medical imaging benchmarks demonstrate that SaFF-AD achieves competitive or superior anomaly detection performance compared to back-propagation-trained models, while requiring substantially fewer parameters and forward evaluations. The forward-forward goodness signal enables self-supervised anomaly and OOD detection without auxiliary networks, post-hoc uncertainty estimation, or heuristically designed scoring functions. These results establish forward-forward learning as a viable and practically attractive alternative to conventional deep learning for safety-critical medical image analysis, particularly in settings characterised by constrained computational budgets, limited labelled data, and distributional uncertainty. By treating anomaly detection as an intrinsic property of the learned model rather than a post-hoc addition, SaFF-AD offers a unified framework that is interpretable, efficient, and well-suited to the open-world conditions encountered in real-world clinical deployment.

Tags

Topics