<p>The meaningful content in a seal image comprises the design elements of the seal, such as characters, iconographic representations, and religious figures. These elements are essential as they provide valuable clues for interpreting the seal. The seal content is often enclosed within a border pattern known as grènetis, and its detection is a useful preliminary step prior to focus on the meaningful content. However, detecting the grènetis on Byzantine seals presents significant challenges due to broken borders. To address this issue, we propose a hybrid AI approach that combines Convolutional Neural Networks (CNNs) with an ellipse estimation method informed by prior knowledge on grènetis shape. In the first step, we utilize a U-Net architecture enhanced with an Atrous Spatial Pyramid Pooling (ASPP) module to effectively capture contextual information in the seal images. By training this model using a loss function that combines Dice and Cross-Entropy losses, mean Dice (mDice), mean Precision (mPrecision), and mean Recall (mRecall) scores of 62.76%, 78.07%, and 53.24% are achieved, respectively. In the second step, we fit an ellipse to the segmented parts of the grènetis, by minimizing a Huber function applied on the Euclidean distance between the fitted ellipse and the segmented grènetis fragments. Our approach yielded a mean Dice score of 92.68%, with a standard deviation of 6.04%, demonstrating the effectiveness of our hybrid method in delineating the meaningful content of the seals. These results make the next stage possible, namely interpreting this content.</p>