Optimising Neural Network Accuracy in Histopathological Image Classification through Advanced Histogram Equalization: A Statistical Approach

Abstract

Histopathological image analysis is critical for accurate disease diagnosis, yet its effectiveness is often hindered by poor image contrast. This study presents an in-depth performance analysis of various histogram equalisation techniques for classifying histopathological images using neural networks. We evaluated global, adaptive, contrast-limited adaptive, bi-histogram, dualistic sub-image, brightness-preserving bi-histogram, dynamic, and Gaussian Mixture Model-based histogram equalisation (HE) on two datasets—binary and multiclass ovarian classification. We trained multiple neural networks with varying complexities, including VGG-16, Resnet-50, and EfficientNetB0, as well as convolutional block attention modules (CBAM) integrated into VGG-16, to analyse data intricacy capture. Using multiple callbacks to handle the learning rate on plateau and early stopping, we ensured high generalisation of models. We statistically assessed performance differences among various HE techniques using ANOVA, observing variations in neural network classification performance across multiple data folds. Results indicated that GMM-HE significantly improved mean classification accuracy by 2.56%, 4.7%, and 6.3% for binary classification and by 0.76%, 3.97%, and 6.4% for multiclass classification with VGG-16, Resnet-50, and EfficientNetB0, respectively. Furthermore, GMM-HE showed a 4.89% improvement on the VGG-CBAM network. Statistical analysis confirmed these performance differences were significant, with GMM-HE consistently outperforming other methods. Our findings highlight the potential of advanced histogram equalisation algorithms to enhance the diagnostic capabilities of neural network models in histopathological image classification. This study gives important information about how well different neural network architectures work with contrast enhancement techniques. This makes it possible for automated histopathological analysis to be more accurate and reliable.