Classificação de lesões cutâneas em imagens infravermelhas com redes neurais convolucionais

Abstract

This work investigates the use of convolutional neural networks for the classification of benign and malignant skin lesions from dynamic infrared thermography images, aiming to support skin cancer diagnosis. Based on the hypothesis that the discriminative signal is associated with the lesion's reheating dynamics rather than with an isolated static frame, each thermal sequence was treated as the primary unit of evaluation. Characterized as applied experimental research, the study used biopsy-confirmed labels, five-fold patientgrouped cross-validation and, as the main metric, the pooled AUC with a 95% confidence interval obtained by stratified bootstrap resampling (also verified with a patient-grouped bootstrap). A custom convolutional network trained from scratch, transfer learning (ResNet18) and classical models (SVM and Random Forest) built on representations extracted from the fine-tuned ResNet18 together with thermal-dynamics features were compared. The best models reached an AUC of approximately 0.66, with no detectable difference from the Random Forest under paired bootstrap analysis, whereas the network trained from scratch showed performance compatible with chance level. None of the tested interventions (multichannel temporal input, feature enrichment, more expressive pooling and model ensembling) surpassed this ceiling. The results suggest that sample size is one of the main constraints on this dataset, although acquisition, radiometric representation and temporal modeling limitations may also have contributed to the observed performance ceiling. The work contributes a reproducible pipeline and a careful assessment of the feasibility of dynamic thermography combined with machine learning as a complementary tool for dermatological diagnosis.