Multi-color space image enhancement with weighted fusion and lightweight deep learning framework for AI-based breast cancer histopathology image classification

Introduction Variations in Hematoxylin and Eosin (H&E) staining, illumination, tissue preparation, and image acquisition affect breast histopathology image analysis, altering nuclear contrast, color consistency, and downstream feature extraction. Separately, prior BreakHis studies have often used image-wise splitting, which can place a patient’s images in both training and test sets and inflate performance estimates. Methods To address these two issues, this study introduced a Multi-color Space Image Enhancement (MSIE) approach and evaluates it under strict, leakage-controlled, patient-wise five-fold cross-validation on the BreakHis dataset. MSIE combines H&E color deconvolution, stain normalization, hematoxylin contrast processing with Contrast Limited Adaptive Histogram Equalization (CLAHE), Hue-Saturation-Value (HSV) space processing, eosin reconstruction, and H&E-weighted fusion. Before model training, the dataset underwent a decoded-pixel Secure Hash Algorithm 256-bit (SHA-256) integrity check, and 125 duplicate images were removed, reducing the dataset from 7,909 to 7,784. The Lightweight Convolutional Neural Network (LCNN) model is trained and tested on MSIE and original RGB images using an identical architecture and patient partitions, and MobileNetV2, ResNet50, DenseNet169, and EfficientNetB0 are evaluated under the same protocol to assess efficiency-accuracy trade-offs across model capacities. Results Performance is assessed at the image and patient levels using accuracy, precision, recall, F1 Score, Receiver Operating Characteristic Area Under the Curve (ROC AUC), and Precision-Recall Area Under the Curve (PR AUC), alongside Fusion-coefficient sensitivity, magnification-specific performance, computational efficiency, Grad-CAM visualization, and independent BACH testing. At the patient level, original RGB achieved an accuracy of 0.8397 ± 0.0325 and F1-score of 0.8789 ± 0.0326, versus a mean accuracy of 0.7522 ± 0.2218 and F1-score of 0.7632 ± 0.2734 for MSIE. MSIE, however, produced a slightly higher mean ROC-AUC (0.9024 ± 0.1274 versus 0.8952 ± 0.1023). Higher-capacity ImageNet-pretrained architectures performed better on MSIE inputs, with EfficientNetB0 achieving the highest mean patient-level accuracy (0.9007 ± 0.0718) and ROC-AUC (0.9707 ± 0.0354). Independent evaluation on the BACH dataset using frozen BreaKHis-trained MSIE-LCNN models yielded an accuracy of 0.7056 and an ROC-AUC of 0.7012, indicating substantial cross-dataset domain shift. Discussion Overall, MSIE provides a stain-aware representation whose downstream benefits are not universal and depend on the classifier architecture, fusion balance, operating point, and dataset domain, thereby supporting leakage-controlled and externally validated assessment of histopathology processing methods rather than claims of immediate clinical translation.