A Patch-Based Computational Framework for the Analysis of Structurally Heterogeneous Bioelectrographic Images

Image datasets characterized by high intra-image structural heterogeneity pose significant challenges for supervised classification, particularly when local patterns contribute unevenly to image-level decisions. In such scenarios, direct image-level learning may obscure relevant local variability and introduce bias in both training and evaluation. This study proposes a statistically guided, patch-based computational pipeline for the automatic classification of elementary morphological patterns, with application to bioelectrographic imaging data. The pipeline is progressively refined through explicit statistical diagnostics, including image-level data splitting to prevent data leakage, class imbalance handling, and decision threshold calibration based on validation performance. To further control structural bias across images, a continuous image-level descriptor, denoted as \textit{pct\_point\_true}, is introduced to quantify the proportion of point-like structures and support dataset stratification and stability analysis. Experimental results demonstrate consistent and robust patch-level performance, together with coherent behavior under complementary image-level aggregation analysis. Rather than emphasizing architectural novelty, the study prioritizes methodological rigor and evaluation validity, providing a transferable framework for patch-based analysis of structurally heterogeneous image datasets in applied computer vision contexts.