Evaluation of fractal descriptors, deep features and XAI representations with LASSO-regularized Hermite polynomial classifier for H&E histological image classification

Author(s)

Pereira, Danilo C. ; Silva, Adriano B. ; Longo, Leonardo H. C. ; Loyola, Adriano M. ; Cardoso, Sergio, V ; de Faria, Paulo R. ; Tosta, Thaina A. A. ; Neves, Leandro A. ; Martins, Alessandro S. ; Nascimento, Marcelo Z.

• Document type: Journal article
• Source: Biomedical Signal Processing and Control; v. 112, p. 18-pg., 2025-08-07.

Abstract

Diagnostic support systems play an important role as a supplementary tool for the identification and recognition of patterns in histological images. This study investigates and compares classification approaches for H&E histological images. The performance was evaluated with a Hermite polynomial (HP) classifier combined with LASSO algorithm and distinct feature datasets derived from: (i) fractal geometry (fractal dimension, lacunarity, percolation), (ii) deep learning features extracted from CNNs (ResNet-50, EfficientNet-B2), and (iii) fractal descriptors applied to explainable AI (XAI) representations (Grad-CAM, LIME). The best features were selected with LASSO regularization and classified separately using the HP algorithm. These experiments employ the 10-fold cross-validation technique. The proposed approach, which is the HP classifier combined with LASSO-regularized deep features (ResNet-50), demonstrated the highest performance, providing accuracy values of approximately 99% and an imbalance accuracy metric of 0.99 for the investigated image datasets. These results were compared with other machine learning algorithms and literature findings. The study highlights the effectiveness of the regularized HP algorithm, particularly when combined with deep features, for classifying histological lesions. (AU)

• FAPESP's process: 22/03020-1 - Normalization of H&E stain by autoencoders with analysis of ensemble learning for histological images
• Grantee: Thaína Aparecida Azevedo Tosta
• Support Opportunities: Research Grants - Initial Project