IMPROVING MASK R-CNN FOR THE DETECTION AND SEGMENTATION OF BREAST TUMORS

ABSTRACT

Detecting and segmenting breast cancer is a crucial endeavor in analyzing medical images, holding considerable importance for identifying and treating the condition early. This study aims to improve the detection and segmentation of breast tumors using the Mask R-CNN (Region-based Convolutional Neural Network) architecture. The research methodology involves comprehensive data collection, annotation, and preprocessing of a dataset containing 891 benign, 421 malignant, and 266 normal breast cases obtained from Kaggle. The Mask R-CNN model was trained and tested on this dataset. The model architecture leverages Region Proposal Networks (RPNs) to identify candidate tumor regions, followed by classification and segmentation branches. The classification branch categorizes the detected regions into benign, malignant, or normal cases, while the segmentation branch generates precise pixel-level masks for individual tumor instances. The results demonstrate the model's strong performance, with high classification accuracy in distinguishing between benign, malignant, and normal breast tissue. The model also achieved good segmentation accuracy, with a mean Intersection over Union (IoU) of 0.75. Evaluation on the training dataset using 10-fold cross-validation yielded a sensitivity of 0.9, specificity of 0.7, overall accuracy of 0.88, and a Dice Similarity Coefficient for segmentation with a mean of 0.82. These results underscore the promise of the Mask R-CNN framework for automating the detection and segmentation of breast tumors, which could greatly assist in early diagnosis, treatment strategizing, and ongoing monitoring of breast cancer. This project emphasizes the necessity for further exploration and progress in this crucial healthcare domain.