DSNet: Automatic Dermoscopic Skin Lesion Segmentation

Automatic segmentation of skin lesion is considered a crucial step in Computer Aided Diagnosis (CAD) for melanoma diagnosis. Despite its significance, skin lesion segmentation remains a challenging task due to their diverse color, texture, and indistinguishable boundaries and forms an open problem. Through this study, we present a new and automatic semantic segmentation network for robust skin lesion segmentation named Dermoscopic Skin Network (DSNet). In order to reduce the number of parameters to make the network lightweight, we used depth-wise separable convolution in lieu of standard convolution to project the learned discriminating features onto the pixel space at different stages of the encoder. Additionally, we implemented U-Net and Fully Convolutional Network (FCN8s) to compare against the proposed DSNet. We evaluate our proposed model on two publicly available datasets, namely ISIC-2017 and PH2. The obtained mean Intersection over Union (mIoU) is 77.5 % and 87.0 % respectively for ISIC-2017 and PH2 datasets which outperformed the ISIC-2017 challenge winner by 1.0 % with respect to mIoU. Our proposed network also outperformed U-Net and FCN8s respectively by 3.6 % and 6.8 % with respect to mIoU on the ISIC-2017 dataset. Our network for skin lesion segmentation outperforms other methods and can provide better segmented masks on two different test datasets which can lead to better performance in melanoma detection. Our trained model along with the source code and predicted masks are made publicly available.

💡 Research Summary

This paper presents DSNet (Dermoscopic Skin Network), a novel and lightweight convolutional neural network architecture designed for the automatic and robust segmentation of skin lesions in dermoscopic images, a critical step in computer-aided diagnosis (CAD) systems for melanoma detection. Recognizing the challenges posed by the high variability in lesion color, texture, and ambiguous boundaries, as well as artifacts like hair and reflections, the authors propose a solution that balances high accuracy with computational efficiency.

The core innovation of DSNet lies in its architecture. The encoder is based on DenseNet-121, which utilizes dense blocks and transition layers. This design promotes feature reuse throughout the network, alleviates the vanishing gradient problem, and leads to a more compact model. The decoder employs depth-wise separable convolutions instead of standard convolutions. This technique significantly reduces the number of trainable parameters, making the network lightweight without sacrificing representational power. Furthermore, inspired by U-Net, DSNet incorporates skip connections that channel-wise concatenate feature maps from the encoder to corresponding layers in the decoder. This mechanism helps recover spatial information lost during down-sampling, enabling precise boundary delineation. The model is trained using a combined loss function of cross-entropy and Intersection over Union (IoU) to optimize both pixel-wise accuracy and region overlap.

The proposed method is rigorously evaluated on two publicly available benchmark datasets: ISIC-2017 and PH2. DSNet achieves a mean Intersection over Union (mIoU) of 77.5% on the ISIC-2017 dataset and 87.0% on the PH2 dataset. These results not only outperform the winner of the ISIC-2017 segmentation challenge by 1.0% in mIoU but also exceed the performance of two well-established baseline models, U-Net and FCN8s, by 3.6% and 6.8% respectively on the ISIC-2017 dataset. The authors emphasize that DSNet accomplishes this state-of-the-art performance while being a parameter-efficient network.

In conclusion, DSNet demonstrates a successful integration of architectural advancements—DenseNet-based encoding for rich feature extraction, depth-wise separable convolutions for efficiency, and skip connections for precise localization—to address the complex problem of skin lesion segmentation. The work highlights the potential for developing accurate yet efficient deep learning models for medical image analysis. The authors have made their trained model, source code, and predicted masks publicly available to foster further research and reproducibility.