From Pixels to Prognosis: Exploring from UNet to Segment Anything in Mammogram Image Processing for Tumor Segmentation

Breast cancer, is the most prevalent form of cancer among women globally accounting for nearly one in four of all new cancer cases. This high prevalence makes breast cancer the second leading cause of death among women making early detection of breast cancer crucial for reducing mortality rates. Mammography, a widely employed medical imaging technique, emerges as a front-line defense against breast cancer. The research aims to achieve pivotal objectives: Firstly, to examine the Segment Anything Model (SAM), created by MetaAI in 2023, and also discuss the model’s use in medical image segmentation. Secondly, the study aims to conduct a comprehensive comparative analysis of SAM’s segmentation capabilities with other existing segmentation models for medical images. The unique strength of the research lies in incorporating both mass and calcification mammograms within the diverse Curated Breast Imaging Subset of DDSM (CBIS-DDSM) dataset this acknowledges the varied nature of breast cancer, exposing the segmentation models to a wide range of scenarios. Through augmenting the dataset with vertical, horizontal, and rotational transformations, this enables accurate identification and isolation of Region of Interests (RoIs) in mammograms, for robust model training for real-world breast cancer diagnosis. The study investigates different segmentation models for medical image segmentation, with an emphasis on prompt-based (SAM), Encoder-Decoder (UNet, UNet++, Attention-UNet, SegNet) models. The study extends to evaluating these models based on Key Performance Indicators (KPIs) such as Dice score, Intersection over Union (IoU) and Accuracy. The insights gained from this research have broader implications for the application of more accurate segmentation in medical image analysis, and making a significant contribution to the continuous efforts to improve breast cancer diagnostic methodologies.