Automatic skin lesion segmentation on dermoscopic images is an essential step in computer-aided diagnosis of melanoma. However, this task is challenging due to significant variations of lesion appearances across different patients.Clinical treatment of skin lesion is primarily dependent on timely detection and delimitation of lesion boundaries for accurate cancerous region localization.
For the accessing the interactive notebooks:
We present a fully automated architecture for accurately detecting border and segmenting the skin lesion by coupling a deep learning model with the Berkeley wavelet transform map derived from the specific kernel filters. Our proposed method effectively combines the Berkeley wavelet transform feature maps into a deep learning U-Net model trained in an end-to-end manner, which results in the reduction of the number of trainable parameters. The model was trained on HAM10000 and ISIC training data and test on PH2 and ISIC Validation data.
The dermoscopic images are pre-processed using Berkeley wavelet transform by applying the 8 mother wavelets on the dermoscopic images to form the berkeley wavelet decomposed image.
| Dataset | Sensitivity | Accuracy | Dice | Jaccard Similarity |
|---|---|---|---|---|
| PH2 | 96.44 | 96.89 | 96.42 | 93.11 |
| ISIC 2017 | 92.46 | 95.64 | 88.20 | 78.89 |
The Use of inexpensive Berkeley wavelet transform helps in enhancing the minutiae details of the skin lesion that helped in achieving an improved algorithm for accurate and automatic skin lesion segmentation. Furthermore, the model is computationally efficient with 1.2 Million parameters, and only takes 0.0625 seconds to segment the lesion from the dermoscopic image, making it extremely useful in clinical settings.