Skin cancer, especially melanoma, has been on the rise in recent years, and it is the leading cause of fatality among all types of skin cancers. Early detection and treatment of melanoma can be effective in remedying the disease.
The risk of death increases with the extent of penetration into the skin, and if melanoma has spread distantly, the five-year survival rate of patients may be as low as 23%. On the other hand, if detected early, more than 90% of melanoma patients survive beyond 5 years. Detecting early-stage melanoma can be challenging even for experienced dermatologists because it can resemble a common skin lesion or mole. There is still a considerable risk of missing a melanoma, even when a large number of benign lesions are excised in each patient, especially in individuals who are at high risk and have multiple atypical moles. Therefore, it is crucial to detect the disease at an early stage to initiate prompt treatment before it spreads locally or metastasizes. Our objective is to detect relevant skin lesions that are present on the skin from the images obtained by iToBoS scanner, with a future plan of classifying individual moles and investigating their size.
iToBoS exploits the data from a 3D skin scanner that allows image acquisition of the patients' entire skin. Medical practitioners use this technique to try to find individual moles in each image obtained from complete body scans. A Deep Neural Network (DNN) is being built and trained particularly to detect and segment moles in the captured images to detect all relevant pigmented skin lesions. For this purpose, we are testing a solution based on the DeepLab architecture. DeepLab is a deep learning model developed by Google Research for assigning a label to each pixel in an image through semantic image segmentation. It uses dilated convolutions to increase the receptive field of the network without losing resolution and a multi-scale approach to capture both fine-grained details and global context in the image, leading to more accurate segmentation results. DeepLab has been used in various applications such as self-driving cars, image recognition, and medical imaging and has contributed to advancements in computer vision and image processing. Although the 3D skin scanner captures very high-resolution skin images (24Mpix), they are broken into smaller tiles to preserve privacy. Each tile may include several lesions or no lesions at all. The DNN produces a set of detections of pigmented skin lesions. DNN technology can aid in early detection, even when it mimics a typical skin lesion or mole. This can significantly increase the chances of survival for melanoma patients.
Although detecting skin lesions is a crucial step in the diagnosis and treatment of skin cancer, there are various challenges that come with this process. One of the major challenges is the misclassification of sun damage as pigmented skin lesion. If a patient has a lot of sun damage on the skin, then these sunburns may be confused by the model, being detected as nevus. Another challenge is the overlapping of skin patches in the high-resolution images obtained from complete body scans. This can result in the same lesion being detected multiple times, as it may be present in overlapping tiles. Additionally, skin patches that are not perpendicular to the optical axis of the camera can cause inaccuracies in the shape and size of the lesion, leading to the need for additional tiles to capture the skin patch from a more orthogonal perspective. Other skin-related issues such as birthmarks, moles, stretched skin, freckles, black spots, excessive hair, or tattoos present on the skin patch can also cause false positives or false negatives in lesion detection. To address this issue, it is crucial to include as many samples containing such skin patches as possible to train the DNN model to be more resilient.
In conclusion, detecting skin lesions is a complex process that requires addressing various challenges related to image segmentation and classification. By identifying these challenges and implementing strategies to overcome them, we can improve the accuracy and effectiveness of our skin lesion detection systems and ultimately contribute to better outcomes for patients with skin cancer.
References:
Akdeniz et al., Prevalence and associated factors of skin cancer in aged nursing home residents: A multicenter prevalence study. PLoS one, 14(4), 2019.
Melanoma Research Alliance, 2019.
Balch et al., Final version of 2009 AJCC melanoma staging and classification. J Clin Oncol 27(36), 2009.
Petty et al., Meta-analysis of number needed to treat for diagnosis of melanoma by clinical setting. J Am Acad Dermatol. S0190-9622(20), 2020
Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2018). DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs. IEEE transactions on pattern analysis and machine intelligence, 40(4), 834-848.