Learning-Based Design of Mismatched Filters via Unsupervised Deep Optimization for Coded Excitation Ultrasound

Abstract

Coded excitation is widely used in ultrasound imaging to improve the signal-to-noise ratio (SNR), and, in particular, Barker-code-based pulse compression offers both high axial resolution and SNR gain. However, matched filtering yields elevated sidelobes that can degrade image quality, motivating the use of mismatched filters. Conventional mismatched filter designs rely on numerical optimization (e.g., minimizing integrated sidelobe level (ISL) or peak sidelobe level (PSL) under energy constraints) but are limited by fixed objectives, which restrict design flexibility. This work proposes a deep-learning-based, filter-to-filter framework that takes the transmit waveform as input and directly predicts mismatched filter coefficients via a multi-objective loss. The framework also integrates the transducer impulse response into the design process, enabling more precise mainlobe control and improved sidelobe suppression compared with a matched filter. Because the loss terms can be readily returned to different design goals, thus providing a flexible path for filter design in coded-excitation ultrasound systems.