Segmentation of Skin Cancer Images Using Fully Convolutional DenseNet - Tiramisu (One Hundred Layers)

Authors

  • Primatua Sitompul Universitas Potensi Utama
  • Rika Rosnelly Universitas Potensi Utama
  • Wanayumini Wanayumini Universitas Potensi Utama

Keywords:

Skin Cancer, Segmentation, CNN, FC-DenseNet, Tiramisu

Abstract

The success of identification of skin cancer images based on Computer Aided System (CAD) cannot be separated from the results of segmentation of skin lesions. The segmentation of cancer images automatically becomes a challenge because of the large variety of skin colors, irregular shapes, and the boundary between the skin and the lesion that is at least clear or irregular. The presence of artifacts such as hair, bubbles, rulers, and black blocks contained in the dermoscopy image is very disturbing and requires additional methods to perform a skin recognition system. In this study, an experimental segmentation of the skin lesion area on skin cancer dermoscopy images was carried out using three Fully Convolutional DenseNet – Tiramisu (One Hundred Layers) architectural models, namely FC-DenseNet 103, FC-DenseNet 67, and FC-DenseNet 56 against the HAM10000 dataset. From the third model used, testing with 100 epochs against FC-DenseNet 103 got the highest results from the other models. The evaluation was carried out with the metric parameter Intersection of Union (IoU architecture FC-DenseNet 103, FC-DenseNet 67 and FC-DenseNet 56 respectively got a value of 83.85%, an accuracy of 81.85% and 76.77%. Meanwhile, evaluation using the Dice Coefficient metric parameter, FC-DenseNet 103, FC-DenseNet 67 and FC-DenseNet 56 architectures, respectively, get values of 91.09%, 89.88%, 88.59%. From the test results, it is known that the number of layers in the FC-DenseNet - Tiramisu model affects the segmentation results

Downloads

Published

30-09-2023

How to Cite

Sitompul, P., Rosnelly, R., & Wanayumini, W. (2023). Segmentation of Skin Cancer Images Using Fully Convolutional DenseNet - Tiramisu (One Hundred Layers) . Cendana International Conference on Social and Technology, 1(1), 130–138. Retrieved from http://prosiding.politeknikcendana.ac.id/index.php/CICoST/article/view/381