Dados do Trabalho

Título

ARTIFICIAL INTELLIGENCE FOR MELANOMA SCREENING: DEVELOPING A SMARTPHONE APP

Introdução

Melanoma is a complex and heterogeneous cancer originating from melanocytes and most commonly found in its cutaneous form. Despite the increasing incidence and high lethality due to the potential for metastasis, it has a good prognosis if detected at an early stage. In current medical practice, suspicious lesions are evaluated visually, with the naked eye, or with the use of a dermatoscope. The diagnosis can be confirmed with a biopsy, histology, and specialized pathological interpretation. To better differentiate between harmless and malignant lesions, techniques to improve diagnostic accuracy can be used at each step, making the process faster, cheaper, accessible, and non-invasive.

Objetivo

In this study, we propose the development of an embedded smartphone solution to assist physicians in detecting melanoma from a photo of the patient lesion.

Método

We used the convolutional neural network (CNN) machine learning method to create the model and Google's TensorFlow30 deep learning framework to train, validate and test our network. For application development, Android Studio and the Kotlin programming language were the main choices. The set of images for the database comes from an open-access repository, the ISIC Archive, which is strictly composed of biopsy-proven melanocytic lesions indicated as malignant or benign. As the final classification will be binary, a comparative base of the application contains 1072 melanoma and 3315 non-melanoma, in a total of 4387 images. Moreover, the app has a validation set containing 976 melanoma images and 2835 non-melanoma images.

Resultado

The confusion matrix indicates that the overall accuracy of our system in training set classification was 95.9% (913 correct classification from a total of 976). In particular, the misclassification rates were 2.2% for other skin lesions (143 patches of 2835) and 4.8% for melanoma (22 patches of 976). Sensitivity, specificity, and F1 scores were 97.8, 95.2, and 92.2%, respectively, and Cohen’s kappa was 0.894.

Conclusão

The results show that the development of a low-cost screening method, as our app, can not only prevent misdiagnosis or unnecessary interventions but improve accuracy on places that lack expert doctors, transforming the diagnostic pathways for patients and health services.

Palavras-chave

Skin cancer screening, melanoma, machine learning.

Área

Oncologia - Prevenção, rastreamento e diagnóstico