This article examines the utility of a deep learning (DL) approach to improve segmentation of standard wrist radiographs and enhance scaphoid fracture detection, including occult scaphoid fractures which are not visible on an X-ray. A retrospective study was performed on 1,011 radiographs obtained from 410 wrists, each analyzed with a CT, cone-beam CT, or MRI scan to determine the ground truth. The radiographs were further filtered, with 660 determined to be apparent scaphoid fractures, 58 as occult fractures, and 293 as controls. The DL-model was designed with two layers, the first with the goal of segmenting the scaphoid bone, and the second to search for the scaphoid fracture. Training of the model was then performed using a hand surgeon’s manual segmentation of the scaphoid bone and fracture with the goal of enhancing recognition of subtle fracture-associated features. In the test set, the model demonstrated a sensitivity of 0.86, specificity of 0.83, accuracy of 0.85, and AUC of 0.92, similar in performance to 3 experts (1 musculoskeletal radiologist and 2 hand surgeons; interrater agreement = 0.69) who independently evaluated the test set. Notably, the DL-model detected 41% of occult fractures compared to the experts who detected only 6.8% - 13.7% of the occult fractures. Annotations in fracture areas compared to labeling in previous studies were highlighted as key in improving algorithm precision, and visualizations of the model segmentations were noted to enhance the clinical workflow. Furthermore, while the negative predicted value (the likelihood of a patient not having a fracture) of 0.93 was likely too low for clinical benefit, it could be useful in cases of low clinical suspicion of fractures and to flag down occult fractures for further imaging.