Change detection and melanoma diagnosis

Skin cancer is the most common type of cancer, and early detection is crucial for successful treatment. One approach to detecting skin cancer is to use change detection, where changes in the skin over time are analyzed to identify potential malignancies.

Among clinical guidelines for melanoma screening, short-term monitoring of lesion changes is widely accepted. A melanocytotic lesion requires the dermatologist attention when there is a significant change within three months, and this could lead to excision of the lesion. Individual clinicians' experiences and biases heavily influence the decision to change or not to change, which is subjective. In this case, having an intelligent automatic change detection algorithm would reduce the bias and increase the accuracy.

Traditionally, there are two steps involved in detecting changes in remote sensing images: 1) obtaining a similarity-feature map (e.g., difference image) between a pair of images with various arithmetical operations (e.g., differencing and rotation), and 2) labeling the pixels of the similarity-feature map as changed or unchanged. However, over the past years, deep learning algorithms have become a trend for change detection. One of the most popular deep learning algorithms for comparing two images are Siamese convolutional networks that have been shown to be effective in change detection in skin cancer.

Siamese networks are neural networks constructed from two identical subnetworks that each receive an image as input. The similarity between the two input images is then assessed by comparing the outputs of the two subnetworks. Two photos of the same patch of skin, obtained at various intervals, are input into the network. The network would then evaluate whether there is a noticeable difference between the two images, pointing out potential changes in the skin.

One advantage of using Siamese networks for change detection in skin cancer is that they can handle large variations in image appearance due to changes in lighting, angle, and other factors. This is important because skin cancer can appear differently depending on the individual and the stage of the disease. Siamese networks also have the ability to learn from a large dataset of images, allowing them to improve over time and become more robust in detecting changes.

It should be noted that Siamese networks have been used in several studies to detect changes in skin cancer. A study by (Zhang, 2021) [1] used a unique Siamese structure-based deep network to determine if a lesion had changed. The lesion change detection task was defined as a task assessing the similarity between two dermoscopy images collected for a same lesion in a short period of time. An internal dataset of 1.000 pairs of lesion photographs captured at a clinical melanoma center was created to test the proposed approach. Experimental results on this dataset showed promise for the suggested approach in detecting the short-term lesion change for objective melanoma screening.

In summary, Siamese networks have demonstrated to be a reliable technique for skin cancer change detection. They have shown the ability to recognize changes in pigmented skin lesions and melanoma and are capable of handling significant fluctuations in image appearance. Siamese networks have a large potential as an important tool for early detection of skin cancer, but further study is required to increase their accuracy, efficiency and robustness in determining changes in skin cancer.

 

[1] Zhang, Boyan and Wang, Zhiyong and Gao, Junbin and Rutjes, Chantal and Nufer, Kaitlin and Tao, Dacheng and Feng, David Dagan and Menzies, Scott W. "Short-Term Lesion Change Detection for Melanoma Screening With Novel Siamese Neural Network". IEEE Transactions on Medical Imaging, 2021.