A Hybrid Method for Robust Malignancy-Benignancy Prediction in Mammograms: Seam Carving and U-Net for CNNs

Mammographies are the primary tool for the early diagnosis of breast cancer. While convolutional neural networks (CNNs) demonstrate significant potential in classifying mammograms, preserving relevant image features for accurate segmentation of the regions of interest (ROI) remains a concern. We propose a hybrid approach combining Seam Carving Algorithm (SCA) for content-aware resizing and U-Net segmentation for preprocessing mammograms before classification by CNN models. The workflow was downsizing mammograms while retaining relevant features, then applying U-Net on downsized images for tissue and lesion segmentation. The resulting segmented images were the input to CNN classifiers. We evaluated DenseNet121, ResNet50, and VGG16 Deep Learning models on both film and digital mammography technologies. We computed the models' performances in Seam Carving then Bilinear Interpolation (BSCBI) image preprocessing images where images were downsized from more than 2000×2000 pixels to 750×500 by SCA then resized to 256×256 by Bilinear Interpolation and in Single Bilinear Interpolation (SBI) preprocessing resizing images from more than 2000×2000 pixels to 256×256. BSCBI image preprocessing significantly improved the classification performances in all the models and in both mammography technologies compared to SBI(>0.98% accuracy and sensitivity). Combining Seam Carving preprocessing and U-Net image segmentation significantly improved malignant case detection in all the CNN models of the benchmark, in both mammography technologies. This hybrid approach of image segmentation prior classification by deep learning networks has potential for improving computer-aided (and potentially fully automated) breast cancer detection systems.