Breaking The Black Box: Heatmap-Driven Transparency To Breast Cancer Detection With Efficientnet And Grad CAM

Abstract

Breast cancer is a significant global health concern, with the manual diagnostic process being time-consuming. The introduction of Computer-Aided Diagnosis (CAD) has emerged as a promising solution, facilitating quicker and more accessible assessments. However, concerns persist regarding the trustworthiness of these automated systems, particularly deep learning models, due to their inherently black-box nature. Transparency and interpretability are crucial elements, necessitating methods to visualize and comprehend the decision-making process of the model. This research aimed to enhance the transparency and interpretability of deep learning models for breast cancer diagnosis. The focus was on developing a method to highlight prominent areas of histopathology slides using heatmaps. The "Histopathology Cancer Detection (HCD)" dataset was used in the investigation. Eight EfficientNet models were examined for fine-tuning, and a feature extractor for binary classification. The optimized model is further utilized to get the output of any particular layer or block of the model with GradCAM (Gradient-weighted Class Activation Mapping). Heatmaps are produced to show the area of the picture that contributed most to the classification. Notably, the model architecture remained unchanged to maintain diagnostic accuracy, while the introduction of heatmaps aimed to provide additional insights into the decision-making process. To validate the effectiveness of the proposed approach, human validation was conducted. Domain experts were presented with histopathology images along with the model-generated heatmaps. The purpose of the questionnaire was to obtain expert comments on the highlighted regions' alignment without altering the model architecture to preserve the performance of the model. The combination of the EfiicientNetB7 model as a feature extractor with an SVM activation function outperformed and achieved the accuracy and the area under the curve (AUC) of 98.95% and 0.9886, respectively. This research contributes to the ongoing efforts to make deep learning models for breast cancer diagnosis more transparent and trustworthy.