In this paper, we investigate factors affecting the performance of breast cancer histopathological image classification using the BreakHis dataset and MobileNetV3-Large as the backbone feature extractor.
First, we examine two sources of data leakage: the presence of multi-label patients and image-level dataset partitioning. We evaluate three experimental settings: (a) multi-label patients with image-level partitioning, (b) removal of the duplicate image copies associated with the multi-label patient while maintaining image-level partitioning, and (c) leakage-free patient-wise partitioning. Compared with setting (a), setting (b) resulted in a performance decline of 4-12 percentage points across evaluation metrics, while setting (c) led to substantially larger declines, with the largest decrease observed in specificity. These findings demonstrate that data leakage can substantially inflate performance estimates.
Second, we evaluate several transfer-learning strategies and observe that fine-tuning the middle and late layers of the ImageNet-pretrained network provides the best overall trade-off among the evaluated metrics, achieving an image-level accuracy of 0.60, an F1-score of 0.72, and a sensitivity of 0.80.
Finally, we repeat the leakage-free experiments over ten independent runs using different random seeds to assess the stability and statistical robustness of the results and compare the original GAP-based classification head with a Flatten-based alternative. The mean, standard deviation, 95% confidence interval, and statistical significance of the results were evaluated using performance metrics. Although the Flatten-based classifier achieved higher mean image accuracy, mean patient accuracy, mean sensitivity, and mean F1-score, the differences between the two classification heads were not statistically significant according to the paired Wilcoxon signed-rank test.