The paper deals with the automatic segmentation of multiple objects in a dynamic X-ray radiograph of the knee joint, which shows the patella, femur, tibia and patellar tendon. It proposes a two-level weighted cross-entropy loss function that balances the losses between these objects and improves the performance of the segmentation model over the traditional weighted function. Two comprehensive evaluation metrics were developed for dimensionality reduction of existing metrics and comprehensive evaluation of multi-object models. Based on them, a new scoring criterion for choosing the optimal model was created. The optimal model combines the DeepLabV3+R50c network with a mixed loss function (τ1LCE2 + τ2LDICE + τ3LBD, ratio τ1:τ2:τ3 = 0.50:0.25:0.25). It achieved average values of: 0.8921 (Metric Metric), Cube 0.9373, Accuracy 0.9316, Value 0.9490, HD95 2.9145 and ASSD 1.0309. This model helps radiologists in diagnosis and reduces their workload.