A Comparative Study of Image Segmentation Techniques in Mammograms (Otsu Thresholding, FCM, and U-Net)
Hanifah Rahmi Fajrin, Young Kim, Se Dong Min · 2024
This study aims to employ three image segmentation techniques—Thresholding (Otsu's method), Fuzzy C-Means (FCM) clustering, and the U-Net deep learning model-on mammographic images. By assessing these methods, we seek to identify the most effective approach for different types of mammographic data and provide insights into their practical applications in clinical settings. Pre-processing steps included image conversion, grayscale transformation, median filtering, histogram equalization, and region of interest (RoI) cropping based on given coordinates. Each method's performance was evaluated using the Dice Similarity Coefficient (DSC). The results showed that U-Net achieved the highest accuracy, with DSC scores of 0.84 for benign and 0.81 for malignant images. These findings suggest that U-Net provides the most accurate segmentation, closely matching the ground truth annotations. Additionally, U-Net's ability to learn and represent complex features, its robustness to variations in image quality. However, U-Net is computationally intensive and requires substantial computational resources for training and inference. Followed by Otsu's method, while simpler and effective for images with clear intensity separations, is limited by global thresholding approach, which may not perform well on images with subtle variations despite noise reduction by median filtering. FCM, although capable of handling overlapping intensities, requires careful parameter tuning to achieve optimal results. This study underscores the importance of selecting and optimizing segmentation methods based on specific imaging data characteristics and preprocessing techniques, ultimately contributing to more reliable and accurate breast cancer detection (benign and malignant).