A Lightweight Optimized CNN Framework for Breast Cancer Screening and Subtype Classification

Breast cancer remains one of the most serious diseases threatening human health globally. To address diagnostic subjectivity and insufficient classification granularity in pathological screening, this study proposes a lightweight complementary convolutional neural network (CNN) framework for two distinct yet clinically related tasks. The first model performs binary classification of invasive ductal carcinoma (IDC) (IDC-positive vs. IDC-negative), while the second model conducts fine-grained multiclass classification of four malignant subtypes. Notably, primary patch-level evaluation for the IDC binary task may produce optimistically biased performance due to data leakage, where patches from the same patient can appear in both the training and test sets. The IDC_regular_ps50_idx5 dataset embeds patient/case identifiers within patch filenames, enabling patient-disjoint partitioning to mitigate this risk. On public patch-level test data, the framework achieves a patch-level accuracy of 0.8732 and an AUC of 0.9245, demonstrating superior resource efficiency compared with several state-of-the-art models. Computational analysis shows that the proposed model has 0.435 M parameters and 0.145 G FLOPs—33.8× fewer parameters than VGG16 (convolutional-only), 54.0× fewer than ResNet50, and 9.3× fewer than EfficientNet-B0. To quantify the optimistic bias introduced by patch-level evaluation, we first performed supplementary patient-disjoint 5-fold GroupKFold cross-validation on the IDC dataset using filename-embedded patient identifiers. Under this intra-dataset patient-disjoint protocol, it achieves a mean accuracy of 0.8005 and mean AUC of 0.8783, with a moderate performance drop of approximately 7 percentage points relative to patch-level results. We further conducted external cross-dataset validation on BreakHis with strict patient-disjoint 5-fold GroupKFold cross-validation. Under this cross-dataset protocol, the model yields substantially degraded performance (mean accuracy of 0.5417 ± 0.1738 and mean AUC of 0.5930 ± 0.1680), illustrating significant performance drop when generalizing to completely unseen patients under domain shift. A systematic loss function comparison experiment under patch-level configuration demonstrates that Focal Loss (γ = 2.0) achieves the highest overall patch-level performance (Accuracy = 0.8125, AUC = 0.8980) on the binary IDC identification task, outperforming the baseline cross-entropy loss. The results indicate that the proposed complementary framework achieves competitive classification accuracy with a favorable accuracy–efficiency trade-off within the IDC dataset; however, strong cross-dataset generalization cannot be guaranteed, and patient-disjoint evaluation on multi-institution matched cohorts is required for real-world clinical deployment. Limitations in generalization to minority classes remain. Future work will focus on constructing multi-institution patient-annotated histopathology cohorts for patient-disjoint evaluation and domain-adaptation techniques to mitigate cross-dataset domain shift.