Purpose: Cancer constitutes a frightening and pervasive threat to human health worldwide, and early detection is vital for enhancing treatment outcomes and saving lives, which demands advancements in early detection methodologies. A significant challenge in dermoscopic image analysis is the presence of hair artifacts, which can obscure skin lesions and hinder accurate segmentation. Methods: This study proposed an advanced approach that combines preprocessing for hair removal with state-of-the-art Transformer-based segmentation techniques to enhance the precision of skin lesion detection. Our approach involves a two-step process. Initially, we apply a classical U-Net model specifically trained to remove hair artefacts from dermoscopic images. Then the preprocessed images are fed into the (improved) i-TransAttUnet model, a Transformer-based architecture designed for high-precision segmentation. This model employs multi-level attention mechanisms to capture detailed and contextual information, addressing the limitations of traditional convolution models in handling complex image features with improvement using L2-Regularisation to overcome overfitting. Results: This preprocessing step clears the images from unwanted obstructions, thereby improving the visibility of skin lesions. The proposed method has achieved Dice coefficients of 88.17 \(\,\pm \,\) 0.05 and 85.05 \(\,\pm \,\) 0.03 for datasets ISIC-2017 and ISIC-2018, respectively. Also gained the highest accuracy and Jaccard index with an optimum loss compared to other Unet model variants. The significant decrease in loss from 0.2244 \(\,\pm \,\) 0.02 to 0.2094 \(\,\pm \,\) 0.01 for the dataset ISIC-2018 over the TransAttUnet model reflects the impact of L2-Regularisation on overfitting. Conclusion: The authors have conducted comprehensive evaluations to validate the potential of our method. The results show that our combined approach of hair removal and improved transformer-based segmentation significantly enhances the accuracy of lesion detection.