Patient-wise, Custom Convolutional Neural Network For Cancerous Cell Detection

Cancer detection using histopathological images is a critical area of medical imaging research. While most existing studies adopt an image-wise experimental setup, this approach may lead to unrealistic performance estimates, as images from the same patient can appear in both training and testing sets. To address this limitation, this work proposes a custom Convolutional Neural Network (CNN) model evaluated under a more reliable patient-wise experimental setup. Specifically, a GroupWise K-Fold Cross-Validation strategy is employed to ensure strict patient-level separation between training and testing data. To support reproducibility and to facilitate researchers, a patient-wise version of the BreakHis dataset has been released online at BreakHis_Patient-Wise. Experimental evaluations on this dataset demonstrate the effectiveness of the proposed approach, It achieves an average recall of 99.07%, F1-score of 85.37%, and highest accuracy of 83.95% at 200X. The results show that the model is highly sensitive to malignant cases, favoring clinical safety, although at the cost of increased false positives. This work promotes patient-wise evaluation as a reliable standard for histopathological image classification.