Chest X-ray (CXR) imaging plays a vital role in diagnosing pulmonary tuberculosis (TB), particularly in resource-constrained environments. This study evaluates the impact of several image enhancement methods on the segmentation accuracy of lung fields in TB chest X-rays using the U-Net architecture. The enhancement techniques compared include traditional filters (Frost, Lee, Perona), contrast-limited adaptive histogram equalization (CLAHE) with various parameter settings, a hybrid speckle noise reduction (HSNR) method, and DnCNN, a deep convolutional neural network for image denoising. Combinations of DnCNN with traditional filters were also explored. The segmentation performance was assessed using accuracy, precision, true positive rate (TPR), true negative rate (TNR), and Dice similarity coefficient. Among all tested methods, CLAHE with parameters (clip limit 0.1; grid size 8×8; 128 bins) achieved the best overall performance, with a Dice score of 0.9586, followed closely by HSNR with 0.9505. In contrast, hybrid combinations involving DnCNN showed lower results across most metrics. These findings emphasize that the choice of enhancement method, including deep learning based preprocessing, significantly affects segmentation outcomes and that carefully selected preprocessing can enhance the robustness of deep learning based TB screening systems. © 2025 IEEE.