Answers are grounded in this paper and cite their sources.
Prostate cancer remains one of the most prevalent malignancies among men worldwide, and accurate differentiation between malignant and benign tissue continues to pose diagnostic challenges, particularly when relying solely on conventional histopathology. In this study, we present a label‑free, high‑resolution imaging and machine learning pipeline for prostate tissue classification using two‑photon microscopy and tissue microarray (TMA) samples. Two-photon fluorescence (TPF) images were analyzed from a quality-screened dataset comprising 295 images (146 normal and 149 tumor) representing 135 tissue sample groups (71 normal and 64 tumor). Three approaches were investigated: (a) handcrafted texture descriptors derived from gray-level co-occurrence matrices (GLCM) classified using logistic regression; (b) MobileNetV3-Small-based deep learning architectures comprising a Global Feature Network (GFN), Global–Local Feature Network (GLFN), and Global–Local Ensemble Network (GLEN); and (c) principal component analysis (PCA)-based variants of these deep learning architectures. Model performance was assessed using group-disjoint five-fold cross-validation. Sample-group-level analysis identified lower homogeneity, energy, and angular second moment in tumor tissues after Bonferroni correction for six feature comparisons. The GLCM-based classifier achieved an accuracy of 71.86% and an area under the receiver operating characteristic curve (AUC) of 0.766. In comparison, GFN, GLFN, and GLEN achieved accuracies of 85.42%, 82.71%, and 86.10%, respectively. GLEN demonstrated an AUC of approximately 0.914, with normal and tumor recall values of 84.25% and 87.92%, respectively. Incorporating PCA preserved the classification accuracy of GLEN at 86.10%, suggesting that dimensionality reduction can retain discriminative information within deep feature representations. Overall, the findings demonstrate the potential of combining label-free TPF microscopy with global-local deep learning representations for automated prostate tissue classification. Nevertheless, further evaluation using larger, independent cohorts and verified patient-level validation is necessary to establish generalizability and clinical applicability.
No takes yet. Share an insight, caveat, or question.
Makkithaya et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: