Distorted born iterative method reconstruction in high-noise environments using KNN-based machine learning denoising

Abstract

Ultrasound tomography reconstruction using the distorted born iterative method (DBIM) is sensitive to measurement noise, which degrades image fidelity and slows convergence. We propose integrating a k-nearest neighbors (KNN) denoising step within each DBIM iteration to suppress noise adaptively while preserving structural edges. Simulations with a circular cylindrical target and transmit/receive geometry (12×12) were conducted at signal-to-noise ratio (SNR) levels of 6 dB, 3 dB, and 1 dB. Compared with conventional DBIM employing Tikhonov regularization, the KNN-filtered DBIM reduces normalized reconstruction error by up to 57.2% at 1 dB and shows faster error decay over successive iterations. The method is training-free, computationally lightweight, and preserves fine structural details. These properties make KNN-filtered DBIM attractive for noisy or resource-constrained imaging environments. Future work will validate the approach on experimental data and explore adaptive K selection.