Descrição
Every day, computational vision is growing and being used in health aid systems to
improve performance and reduce the time of several processes. Despite that, the amount
of research on oral lesions segmentation and classification is still very low. Oral and
mouth cancers are the 16th most common form of cancer in the world and are presented
with a high mortality rate when discovered late. One main problem that makes it hard to
detect them in the early stages is the lack of specialized professionals, a gap that can
be minimized by the use of telediagnosis and artificial intelligence. The segmentation
process is already used in dermatology lesions, but there are still few works exploiting
the oral cavity lesions. Such characteristics as borders and asymmetry can assist the
diagnosis of cancer cases, but then a segmentation process is needed. Technologies
such as artificial intelligence and image processing can be used to segment oral lesions,
making the process quicker and allowing the assessment of more cases, thus helping
more people. Of the few studies developed, the ones with the best results used deep
learning to distinguish the lesions. Therefore, this work’s objective is to present and
evaluate different methods for the automatic segmentation of oral macules and stains in
photographic images using pixel-wise intensity features. Three methods to segment oral
lesions were described in this research. They were evaluated in accuracy, precision,
recall, and F1 score. The third method developed had the best performance in the tested
images. It used a backprojection image created from the original inverted grayscale
image and the Otsu binarization in two steps. This method resulted in an accuracy of
0.849, a precision of 0.701, a recall of 0.753, and an F1 score of 0.608. The results
were satisfactory because they achieved values close to the related works, even without
using complex algorithms or artificial intelligence.
