Quantitative measurement of breast cancer features based on histopathological images

Breast cancer is the most prevalent carcinoma in women worldwide. Developing country such as Malaysia is found to have a notable increase in the number of incidence and mortality. The number of cases in 5-year prevalence (i.e., 2014 – 2018) (for all ages) of the breast cancer was 24,021 cases and this number is estimated to undergo persistent deterioration in the near future. Nottingham Histopathological Grading (NHG) system is a semi quantitative grading system for breast cancer. It is accepted as the gold standard in providing overall grade of breast cancer. Mitotic counts, glandular formation, and nucleus pleomorphism are breast cancer features and are considered in the NHG system. Breast cancer grading based on the NHG system is done manually through visual examination under a light microscope to evaluate glandular structures, nucleus features, and mitotic cells. Visual examination on the histopathological slides is a cumbersome and tedious task. The subjective nature of the NHG system may results in intra- and inter-observers variability which can significantly impinge the treatment planning process. This thesis focuses on the image processing techniques to detect mitotic cells, segment tumor regions, and detect glandular structures on the breast cancer images. In the mitotic cells detection, the domain knowledge and strategies of the histopathologists were incorporated to reduce number of false positive (i.e., non-mitotic cells). Glandular formation can be estimated using tumor and glandular areas. The tumor region segmentation started by implementing the Spatial Neighborhood Intensity Constraint (SNIC) to reduce image complexity and followed with the Fuzzy C Mean (FCM) with guided initialization. In the glandular detection, Spatial Angle Index (𝑆𝐴), Distribution Index (𝐷), and Harmonic Mean Index (𝐻𝑀𝜅) were proposed to quantify the glandular features. The glandular detection study comprehensively presents a geometrical viewpoint between the surrounding neoplastic cells and the central lumen. The proposed mitotic cells detection shows promising results with accuracy (𝐴𝑐𝑐), specificity (𝑆𝑝𝑒), receiver operating characteristic (𝑅𝑂𝐶), precision (𝑃𝑟𝑒), recall (𝑅𝑒), and F1 score (𝐹1) of 98.4%, 99.1%, 94.5%, 88.3%, 89.8%, and 89.1%, respectively. In the tumor regions segmentation, the obtained 𝐴𝑐𝑐, 𝑃𝑟𝑒, 𝑅𝑒, 𝐹1, Area Overlap Measure (𝐴𝑂𝑀), and Combined Equal Importance (𝐶𝐸𝐼) are 91.2%, 90.9%, 93.4%, 92.1%, 85.7%, and 90.1% respectively. In the glandular detection, the proposed 𝐻𝑀𝜅 provides 𝐴𝑐𝑐, 𝑃𝑟𝑒, 𝑅𝑒, and 𝐹1 of 97.5%, 95.2%, 100.0%, and 97.6%, respectively when classifying using the testing dataset. The proposed mitotic cells detection and tumor regions segmentation are repeatable and reproducible methods. The segmented tumor regions could be used to calculate the tumor area. The glandular structure is measurable using the proposed 𝑆𝐴, 𝐷, and 𝐻𝑀𝜅. Overall, the findings in this research could be used as a second opinion to histopathologist in facilitating breast cancer grading procedure.