Triagem espectral de staphylococcus spp. em imagens hiperespectrais com yolo26 e aprendizagem positiva-não rotulada

Abstract

This work addresses the problem of building a computational screening pipeline for Staphylococcus spp. samples in hyperspectral images (Hyperspectral Imaging, HSI) acquired in the short wave infrared range (SWIR). Spectral screening is defined as sample prioritization based on similarity between spectral signatures, understood as representative curves of the sample response across the analyzed wavelengths. The experimental dataset contains 19 bacterial samples, including 3 ATCC references and 16 UNKNOWN samples. The proposed pipeline combines radiometric calibration, removal of water absorption spectral windows, Standard Normal Variate (SNV) normalization, pseudo Red, Green and Blue (RGB) image generation, region of interest (ROI) proposal using YOLO26, controlled synthetic positive signature generation, and prototype-based classification with positive-unlabeled learning. The operational output consists of ATCC_LIKE and UNKNOWN calls, defined by a similarity score to the ATCC prototype and by a threshold selected under Leave-One-Image-Out (LOIO) validation. The system detected 170 ROIs, with an average of 8.95 ROIs per image and coverage across all samples. The final threshold was τ = 0.9951, achieving 1.0 sensitivity over the three ATCC references, 0.60 operational precision, 0.75 operational F1 score, and a 12.5% UNKNOWN signaling rate. The prioritized UNKNOWN samples were 503B_240506-160006 and 539_240506-160131. The results indicate that combining HSI in the SWIR range, ROI detection, controlled synthetic generation, and PU modeling is suitable for exploratory spectral screening and for organizing samples by investigation priority.