AI 中文总结
该研究将稀疏MEG源定位建模为QUBO问题，结合残差感知筛选，经模拟退火评估，其定位误差较最优基线MxNE降低62.2%，且可适配量子退火后端。
AI 中文摘要
脑磁图（MEG）源定位是一个不适定的逆问题，因为不同的皮层源配置可产生相似的传感器级场。我们将稀疏多源定位表述为二次无约束二元优化（QUBO）问题，并结合感知残差的候选筛选：从传感器空间残差生成候选源位置组，为这些候选估计固定的传感器空间模板，再通过数据拟合、成对模板交互及软基数项联合选择活动模板。我们在受控合成MEG模拟中采用经典模拟退火算法评估该方法，主要针对双源场景，并与MNE、dSPM、MxNE、LCMV及RAP-MUSIC进行对比。在100次主基准试验中，QUBO的平均感知基数定位误差为8.45 mm，而该指标下表现最佳的基线方法MxNE为22.35 mm，降幅达62.2%。复合指标在按真实源数归一化前，会对每单位源数不匹配施加50 mm的惩罚；由于MxNE在35次试验中仅返回1个源，该降幅同时反映了空间定位和源数性能。在单独的敏感性试验中，QUBO在测试的传感器噪声和源数条件下均保持竞争力，尽管RAP-MUSIC在部分低噪声及三源场景中表现与QUBO相当或更优。本试验采用经典模拟退火算法，未评估量子硬件也未宣称量子优势；所得二元二次目标可直接转换为伊辛（Ising）表示，为后续在量子退火及混合后端上评估提供可能。
英文摘要
Magnetoencephalography (MEG) source localization is an ill-posed inverse problem because distinct cortical source configurations can produce similar sensor-level fields. We formulate sparse multi-source localization as a quadratic unconstrained binary optimization (QUBO) problem combined with residual-aware candidate screening. Candidate source-location groups are generated from the sensor-space residual, fixed sensor-space templates are estimated for the resulting candidates, and active templates are jointly selected using data-fit, pairwise template interactions, and soft-cardinality terms. We evaluate the method using classical simulated annealing in controlled synthetic MEG simulations, primarily under a two-source condition, and compare it with MNE, dSPM, MxNE, LCMV, and RAP-MUSIC. Across 100 main-benchmark trials, QUBO achieved a mean cardinality-aware localization error of 8.45 mm, compared with 22.35 mm for MxNE, the best-performing baseline according to this metric, corresponding to a 62.2% reduction. The composite metric adds a 50 mm penalty per unit of source-count mismatch before normalization by the true source count. Because MxNE returned only one source in 35 trials, the reported reduction reflects both spatial localization and source-count performance. In separate sensitivity experiments, QUBO remained competitive across the tested sensor-noise and source-count conditions, although RAP-MUSIC performed comparably to or better than QUBO in some low-noise and three-source settings. The present experiments use classical simulated annealing and do not evaluate quantum hardware or claim quantum advantage. The resulting binary quadratic objective admits a direct Ising representation, enabling future evaluation on quantum-annealing and hybrid backends.
Comments16 pages, 4 figures
