Convolutional decision trees (CDTs) are machine learning models employed as explicable methods for image segmentation. This is because their graphical structure makes it relatively straightforward to interpret how the tree successively divides the image pixels into two classes, distinguishing between objects of interest and the image’s background. Several techniques have been proposed for the induction of CDTs. However, these techniques often require significant computational time and memory, with some requiring days to complete the induction process. This study proposes two techniques for selecting a representative sample of pixels from an image for the model’s training process: raw selection and median selection. These techniques aim to reduce the computational cost of inducing a CDT while maintaining or improving the resulting segmentation’s precision, measured by the F1-score. The proposed techniques were evaluated using the retinal vessel segmentation database and the SHADE-CDT method to induce a CDT, demonstrating a \(70\%\) reduction in processing time.