A timely and accurate diagnosis is the foundation of efective treatment. In cases in which the diagnosis has to be made based on a complex dataset, such as the one produced by MRI imaging, detecting injuries or other pathologies becomes a time consuming and tedious process susceptible to errors. An example of such a process is the detection of posterior cruciate ligament injuries. The goal of this study is to examine the possibility of automating the detection of such injuries using machine learning methods, in order to facilitate the work of radiologists and reduce the number of misdiagnoses. The study was performed on a sample of 743 PD weighted knee MRI sequences with fat suppression. The sequences were collected at the Clinical Hospital Centre Rijeka, Croatia. A written permission for their use was obtained from the Centre's Ethics Committee. For each sequence the ligament region was marked, and the diagnosis was determined by visual inspection of the sequence images, as well as by consulting the written assessment results. The sequences for which it was not possible to clearly determine the diagnosis were removed from the sample. The automation of injury detection was performed in three stages. During stage one the ROI detection process was implemented. The detection of ROI slices was automated using SVM models, in combination with slice HOG features and labels indicating whether or not the slices are a part of the region of interest. Linear regression models were used to detect the two-dimensional ROI coordinates, in combination with slice HOG features and the vectors of real two-dimensional ROI coordinates for the slices. In stage two the possibilities of injury detection automation were examined, using SVM models, sequence HOG features, and the previously determined diagnoses. During stage three the models were combined in order to evaluate the quality of the final solution. The findings indicate that the models generated to detect the region of interest are, by themselves, efective in detecting MRI slices that display the ligament, as well as the two-dimensional coordinates of the region of interest, but combining the two produces a decline in the quality of detection. The quality of the implemented process of injury detection is unsatisfactory. This finding is probably caused by the insucient injury sample in the dataset. The quality of detection declines further by combining this model with the previous two.