<![CDATA[Acne lesion segmentation is essential for automated dermatological image analysis. However, developing accurate deep learning segmentation models requires pixel-level annotations, which are costly and time-consuming to obtain. To address this limitation, pseudo-mask generation methods can be utilized to produce surrogate labels for training segmentation networks without manual annotation. This study presents a comparative evaluation of five pseudo-mask generation methods for acne lesion segmentation, namely Contour-Based, Superpixel (SLIC), K-Means, Weakly Supervised Semantic Segmentation (WSSS) based on Otsu Thresholding and Morphological Operations, and Pseudo-Mask-Based Generation. The generated pseudo masks were employed as supervisory labels to train a ResNet34-U-Net segmentation model under identical experimental settings. To improve the robustness of the training process, data augmentation was applied exclusively to the training dataset. Segmentation performance was quantitatively evaluated using Dice Score, Intersection over Union (IoU), Precision, Recall and Validation Loss, and qualitatively assessed through visual comparisons of the generated pseudo masks and predicted segmentation results. The experimental results demonstrate that pseudo-mask quality has a substantial impact on segmentation performance. Among the evaluated methods, Superpixel (SLIC) achieved the highest performance with a Dice Score of 0.898, an IoU of 0.815, a Precision of 0.889, a Recall of 0.908, and the lowest validation loss of 0.268 indicating superior lesion boundary preservation and region consistency. WSSS (Otsu + Morphology) also produced competitive results, whereas Pseudo-Mask-Based and Contour-Based methods yielded comparatively lower performance. These findings demonstrate that high-quality pseudo masks can provide effective supervision for ResNet34-U-Net training, offering an annotation-efficient approach for acne lesion segmentation and providing practical insights into selecting suitable pseudo-mask generation methods for dermatological image analysis.]]>
Copyrights © 2026
