Neutron　coded　imaging　is　an　effective　tool　for　diagnosing　the　shape,　size　and　symmetry　of　deuterium　(D)–tritium　(T)　plasma　in　inertial　confinement　fusion　(ICF).　It　can　provide　an　important　reference　for　designing　and　improving　the　DT　pellet　and　confinement　configuration.　Image　reconstruction　algorithms　play　a　role　of　reconstructing　the　source　images　from　the　blurred　coded　images　and　the　point　spread　functions　(PSFs)　of　imaging　systems.　Conventionally,　the　convolution　model　is　used　as　the　mathematical　model　for　neutron　coded　imaging　reconstruction,　but　it　applies　only　to　the　spatially　invariant　PSF.　In　this　paper,　the　linear　equations　model　is　regarded　as　the　mathematical　model　for　the　reconstruction,　and　it　can　also　be　suitable　for　spatially　variant　PSF.　In　the　reconstruction,　the　spatially　variant　PSFs　were　simulated　through　Monte　Carlo　method.　Then　an　improved　genetic　algorithm　(IGA)　for　the　source　image　reconstruction　was　proposed.　The　comparison　of　its　performance　with　other　types　of　deterministic　algorithms　(like　the　algorithm　with　total　variation　(TV)　minimization)　was　conducted,　and　the　results　showed　that　the　IGA　has　better　performance　in　source　reconstruction　regardless　of　the　utilization　of　TV　sparse　prior.　©　2020