Computer Aided Detection in Mammography

Breast cancer is a major health concern being responsible for the second most number of deaths caused by cancer in women. Early detection is therefore crucial to help reduce the number of deaths. At the moment, mammography is the best tool available to screen for early breast cancer. However, the detection of lesions in mammograms is a difficult procedure due to several factors the most prominent one being the high number of non-pathological structures present in the image. The arrival of computers and data networks in the medical environment, besides simplifying the access to patient data and related services, has strongly contributed to improve the quality and decision taking in healthcare. Digital Medical archives and databases are one of the most important improvements that, although still presenting a few limitations, significantly contribute to better medicine at a lower cost. Computer Aided Detection systems are an example of these systems that can help increase the diagnostic efficiency of breast cancer. This thesis was developed under the SPACS project currently ongoing at INESC-Porto and also counted with the orientation in the company of researcher Jakub Tkaczuk (MSc). The project’s aim is to develop a software module to integrate with current Picture Archiving and Communication Systems (PACS) that acts as a CADe and CADx of suspicious lesions in mammograms. The objective of this dissertation is to provide a valuable contribution to the SPACS project mainly in the areas of the evaluation metrics of the detection/segmentation algorithms and also in improving the performance of the algorithms themselves. This work has successfully introduced a new framework to evaluate the performance of microcalcifications or mass detection algorithms that are robust rigorous and predictable. These metrics will hopefully contribute to a better uniformisation of the performance evaluation metrics used within the medical image processing community. The proposed segmentation method obtained some encouraging results because it meant a significant improvement relative to the region growing algorithm since it obtained an additional 20% of overlap area in the enhanced version of the ROI. The implemented microcalcifications approach was compared to the current detection algorithm which highlighted some limitations of both approaches and allowed to devise some directions in the future work.