Multiclass Segmentation of Hepatic Tissues with Emphasis on Peritumoral Region to Support Radiotherapy Planning

Radiotherapy planning can be interpreted as an optimization problem, in which the precise definition of anatomical regions directly influences dose distribution. In this context, medical image segmentation is fundamental, as it provides the main parameters for decision and optimization models.
This work proposes a multiclass segmentation model based on convolutional neural networks (U-Net) applied to computed tomography images, considering not only the liver and tumor, but also a class of adjacent peritumoral tissue. The inclusion of this region aims to capture frequently neglected transition zones, but potentially relevant for therapeutic planning.
The model was trained on data derived from the Liver Tumor Segmentation Challenge, using two-dimensional images in the axial plane. To deal with the imbalance between classes, a weighted loss function was employed. The evaluation was performed using classic segmentation metrics, such as Dice coefficient, precision, recall, and F1-score.
The results indicated high performance for healthy and tumorous liver tissues, with Dice coefficients greater than 0.95. The class of tissues adjacent to the tumor showed greater variability. The model was able to identify consistent spatial patterns in these regions.
From the point of view of Operations Research, the contribution of the work lies in improving the quality of information used as parametric data for optimization models for optimal radiotherapy planning. Explicit modeling of adjacent regions can contribute to a more realistic representation of the zones of interest, favoring more robust approaches in decision support.