Flat-valgus deformity of the foot is a common condition that can lead to various health problems such as pain syndromes and curvature of the spine. For effective diagnosis using software tools, accurate segmentation of the posterior part of the foot in the images is required. In this study, two image segmentation methods were compared: threshold processing and a model based on a convolutional neural network (CNN), namely the U-Net architecture. Threshold processing, although easy to implement, is not always effective on images with uneven brightness or noise. Whereas a neural network-based model is a more complex but more accurate method capable of adapting to different image conditions. The study showed that the neural network-based model demonstrates high accuracy of posterior foot segmentation in images of various patients. The accuracy of this model was 97% on test data and 95% on validation data, which confirms its effectiveness. A convolutional neural network-based model, such as the U-Net architecture, is the preferred choice for image segmentation of the hindfoot. Its ability to adapt to different imaging conditions and high accuracy make it an effective tool in clinical practice.