Análise de arquiteturas de redes neurais para detecção de bacilos em baciloscopia de campo claro usando radiomics

Abstract

Tuberculosis (TB) remains a major global public health concern, particularly in low- and middle-income countries. The most widely used diagnostic method is bright-field sputum smear microscopy, a low-cost technique that, however, depends heavily on the microscopist’s experience, is time-consuming, and is subject to intra- and inter-observer variability. In this context, automated image-analysis approaches have the potential to increase diagnostic accuracy, standardize results, and reduce the workload in clinical environments. Among these approaches, deep learning methods and radiomics-based feature extraction have shown promise in capturing relevant structural and textural patterns. This work presents a comparative analysis of different neural network architectures for bacillus detection in bright-field smear microscopy images, using both raw images and radiomic features extracted with PyRadiomics. The dataset consists of 28,848 negative images and 2,636 positive images, later balanced to 2,635 samples per class. Two radiomics pipelines were evaluated: (i) extraction using segmented masks only for positive images and full masks for negative images, and (ii) extraction using full masks for all images. In both cases, feature selection involved removing zero-variance features, applying correlation-based filtering, and ranking features via Random Forest importance. In the first pipeline, both MLP and Random Forest achieved perfect performance (accuracy, precision, recall, and F1-score equal to 1.00). However, this performance does not reflect the true nature of the problem: it was caused by information leakage, mainly due to features such as diagnostics_Mask-original_VoxelNum, which remained constant for all negative images because of the full mask, thus creating an artificial pattern highly correlated with the class label. In the pipeline using full masks for all images, the performance of Random Forest (0.89) and MLP (0.88) reflected a more realistic evaluation of the radiomic approach. The convolutional neural network trained directly on the raw images achieved an accuracy of 0.99, demonstrating high robustness and serving as a reliable reference for comparison with radiomics-based methods.