Sistema de detecção e classificação multirrótulo de erros de aquisição em imagens de impressões digitais neonatais

Abstract

Newborn identification contributes to clinical traceability, the correct association between the patient and the medical record, and the reduction of inconsistencies in care. Fingerprint recognition is a biometric modality used to identify individuals, but the acquisition of neonatal images presents specific challenges: papillary ridges are narrow, the skin is susceptible to deformation, the finger moves easily, and the sensor may accumulate residues during the procedure. In this context, scalar quality metrics, such as the NIST Fingerprint Image Quality 2 (NFIQ2), are useful for summarizing image quality, but they do not inform the operator of the type of error present or the corresponding corrective action. In this work, a multi-label system for detecting acquisition errors observable in neonatal fingerprint images was developed and evaluated. The system was conceived as a complement to these metrics by providing actionable labels and real-time operational feedback to the operator during acquisition. The method included the formalization of a catalog with nine error or quality labels, the definition of a severity scale, the use of the BenaPRO annotation tool, the preparation of a traceable tabular dataset, the application of consistency rules, partitioning by newborn, and the training of an EfficientNet-B0 classifier with multi-label output. The final labeled dataset used in the development and evaluation of the classifier comprised 793 images, distributed among the training, validation, and test sets. The model was evaluated with label-calibrated thresholds and multi-label classification metrics. On the test set, the classifier achieved a Macro-F1-score of 0.384, a Micro-F1-score of 0.478, and an mAP of 0.460. Performance varied across labels, with limitations in the classes with lower support. The average inference latency was approximately 0.013 s per image on CPU, with an estimated computational cost of 0.385 GMACs for 224 x 224 pixel inputs, indicating temporal compatibility with the proposed use. The results indicate the technical feasibility of the multi-label approach to support the identification of acquisition errors, although application in an operational environment still depends on the integration and evaluation of the system in the acquisition workflow.