Accurate skin lesion segmentation is an important task in dermatology for facilitating early diagnosis and treatment planning. The challenges in skin lesion segmentation comprehend the variability in lesion, low contrast, heterogeneous backgrounds, overlapping or connected lesions, noise and certain artifacts. Despite of these challenges, Deep learning models accomplish remarkable results for skin lesion segmentation by automatically learning discriminative features. The current research introduces a novel approach utilizing the ASSP-based Deeplabv3+ for skin lesion segmentation along with other UNET-based learners while employing VGG-16, VGG-19 and Dense nets as encoders. In addition, an analysis is conducted on GAN-UNET to evaluate the potential of Generative Artificial Intelligence in generating segmented images of skin lesions. Three benchmark medical image datasets, namely ISIC-2016, ISIC-2018, and HAM10000 Lesion Boundary Segmentation, are used to evaluate all five models. The models are trained exclusively on the ISIC-2018 dataset. A comparative analysis is performed, comparing the performance of these models against state-of-the-art segmentation methods, focusing on standard computer vision metrics. The proposed Deeplabv3+ model outperforms by showcasing its ability to accurately delineate skin lesions and surpassing existing techniques in terms of segmentation accuracy as 0.97, Jaccard coefficient as 0.84 and dice coefficient as 0.91.

准确的皮肤病变分割是皮肤病学的一项重要任务，有助于早期诊断和治疗计划。皮肤病变分割的挑战包括病变的变异性、低对比度、异质背景、重叠或连接的病变、噪声和某些伪影。尽管面临这些挑战，深度学习模型通过自动学习判别特征，在皮肤病变分割方面取得了显着的成果。当前的研究引入了一种新颖的方法，利用基于 ASSP 的 Deeplabv3+ 以及其他基于 UNET 的学习器进行皮肤病变分割，同时采用 VGG-16、VGG-19 和 Dense 网络作为编码器。此外，还在 GAN-UNET 上进行了分析，以评估生成人工智能在生成皮肤病变分段图像方面的潜力。三个基准医学图像数据集，即 ISIC-2016、ISIC-2018 和 HAM10000 病变边界分割，用于评估所有五个模型。这些模型仅在 ISIC-2018 数据集上进行训练。进行比较分析，将这些模型的性能与最先进的分割方法进行比较，重点关注标准计算机视觉指标。所提出的 Deeplabv3+ 模型表现出色，展示了其准确描绘皮肤病变的能力，并在分割精度为 0.97、Jaccard 系数为 0.84 和 dice 系数为 0.91 方面超越了现有技术。