Density Dedicated Deep Learning Model for Mammogram Malignancy Classification
Mehdi Amini, Yazdan Salimi, Zahra Mansouri, Hossein Arabi, Isaac Shiri, Habib Zaidi · 2022
A body of literature has reported the promising performance of deep learning models when applied to mammograms for different clinical tasks. However, a major pitfall of deep learning models is when they are applied to mammograms bearing high-density breasts since the overlapping high-dense tissue can cover the lesion and make diagnosing and interpreting it difficult. Thus, analysis of dedicated models trained for low- and high-dense breasts can be of great importance. In this study, we aimed to develop low- and high-density dedicated deep-learning models for classifying breast masses on mammograms into benign and malignant cases. Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM) dataset, including mammograms, cropped to masses, and the pathologic diagnoses, were adopted for this study. For dedicated models, the dataset was split into low-(BI-RADS density1 and 2) and high-dense (BI-RADS density3 and 4) groups. Contrast-limited adaptive histogram equalization (CLAHE) was applied to enhance the contrast of the images, and then all images were resized to 255 × 255 matrix size and normalized. A modified DenseNet-201 deep neural network was trained with the learning rate starting at 0.0001 and decreased in a piecewise manner every epoch with the RMSProp optimizer. Fifteen percent of training data was excluded for validation, and training was continued for 100 epochs. Data augmentation, including rotation, flipping, and scaling, was implemented on the training dataset to prevent overfitting. The trained model was evaluated using the test set. Accuracy (ACC), area under the receiver operating characteristic curve (AUC), sensitivity (SEN), and specificity (SPE) for the general model were 0.720, 0.771, 0.732, 0.701, for the Low-density model were 0.788, 0.818, 0.824, and 0.742, and for High-density model was 0.712, 0.621, 0.962, and 0.180, respectively. Our study highlights the importance of developing dedicated models tailored to the nature of high-density breasts to improve overall accuracy. We also suggest performing specific preprocessing for high-density breasts, such as region-wise contrast enhancement for the regions with high-intensity values (with a gamma filter, etc.).