This thesis outlines two new methods to drastically improve the reconstruc- tion capabilities of neutrino events at Deep Underground Neutrino Experiment (DUNE), in the Pandora reconstruction framework. The liquid argon time projection chamber (LArTPC), the detector technology of choice at DUNE, provides high spatial and calorimetric resolutions, presenting a difficult but exciting reconstruction problem. One of the main reconstruction frameworks for event reconstruction in LArTPCs is Pandora, a software development kit using a multi-algorithm approach to pattern recognition, which is designed to target the complex pattern recognition problems that occur in particle physics. The work in this thesis includes an overhaul of the 3D event reconstruction for tracks, producing 3D hits from combinations of underlying 2D positions. This new method produces more coherent and truthful 3D representations of tracks, by intelligently selecting hits from a generated 3D point cloud through stages of fitting. Secondly, a graph neural network (GNN) is utilised for the complex problem of electromagnetic shower growing, taking an electron or photon shower that is clustered into hundreds of small groups and producing larger, more representative clusters per shower, whilst avoiding contamination from other interactions in the event. Deep learning is used to give a more global view of the event for growing, and to better use the topological features of the showers to help the growing process. All this work is verified on DUNE far detector simulated data, to give an understanding of what performance gains are made, and the failure modes they fix. Verification of the deep learning method is performed on real test beam data from ProtoDUNE Single-Phase (ProtoDUNE-SP) at CERN. This verification helps give confidence that work performed on simulated data can also be applied to real data, which is especially interesting for methods that utilise deep learning.