Automated breast cancer diagnostics using a Digital Image Elasto Tomography (DIET) system.
|Fitzjohn, Jessica Louise
|Breast cancer was the most common cancer in the world in 2020 and is a major worldwide health concern. For women it accounts for a quarter of all cancer cases with high mortality of 680,000 annual deaths in women. Screening programs are necessary to reduce breast cancer mortality by finding cancer at an earlier, more curable stage, increasing treatment options, reducing treatment costs, and improving outcomes. Ideally, screening would equitably cover women of all ages, as 2% and 11% of breast cancer occurs in women under 30 and 40 years of age, respectively. Currently, the only approved large-scale breast screening technology is x-ray mammography, which has a number of issues making it unsuitable as an equitable breast screening tool. Limitations of mammography include painful breast compression, harmful radiation exposure and poorer performance in younger women and women with dense breast tissue (50% of women). Small radio-density contrast results in radiologist-dependent performance and a high false positive rate. In addition, mammography is expensive and requires infrastructure, such as x-ray shielded rooms, making it unsuitable for screening programs in developing countries and much less accessible for many women who live rurally. These issues contribute to a lower screening participation, as well as no publicly funded screening solution for younger women, which leads to cancer being found at later stages with consequently worse outcomes. This thesis presents a breast screening technology capable of overcoming these screening limitations and develops clinically feasible, automated diagnostic algorithms to be used in con- junction with this technology to provide higher diagnostic accuracy than mammography. First, the need for a different breast screening technology is identified by analysing the 2004 Kew Zealand change in screening eligibility age and using this analysis to quantify the socio-economic benefits of providing a more equitable breast screening solution. This analysis clearly showed the high number of women found with larger tumors when they initially enter into screening, where previously cancer has grown unchecked. This incidence of larger tumors highlights inequity based on age, where younger women contracting breast cancer are more likely to have worse outcomes and lower survival than those eligible for screening. Ethnic inequities also exist with M-aori women, the indigenous people of Kew Zealand, who are 21% more likely to be diagnosed with breast cancer than non-M-aori, and 68% more likely to die from it. Worse outcomes for M-aori are likely a result of lower breast screening participation due to a large portion of M-aori living rurally, away from main centres were mammography is exclusively provided, which demonstrates inequity based on access. Further, a 'diffi cult' history of poor interaction between M-aori and health providers means some may skip screening and/or experience inequity due to current or post racial bias. Quantification of socio-economic benefits showed increasing screening age eligibility to all women 20+ could potentially save 43 lives (∼7%) annually and reduce treatment costs by 10.1%. Increasing screening participation for the currently screened age group (45-69) to 90% could save 37 lives (∼5.5%) annually and reduce costs by 14.5%. Implementing a screening device with the capacity for both increased screening participation and expansion of screening age eligibility could save 102 lives (∼16%) annually and result in a 33.1% reduction in breast cancer treatment costs. Thus, significant social and economic benefits could be realised with the use of a new screening technology capable of overcoming mammography's limitations, such as increasing access and providing safe screening for all women. Second, this thesis suggests a technology capable of providing these socio-economic benefits and critically analyses the diagnostic accuracy of mammography to generate diagnostic criteria to assess diagnostic algorithms. Digital Image Elasto Tomography (DIET) has been developed as a breast screening technology to overcome limitations of mammography. The technology is portable, low-cost and with no requirement for additional infrastructure, making it suitable for use in any clinic and giving it the ability to increase access to breast screening for all women. Further, DIET testing is non-invasive with no harmful radiation exposure and is thus suitable for women of all ages. Breast screening using DIET is comfortable involving a women lying face down, with low-amplitude steady state vibration of one free hanging breast. Surrounding cameras capture images of breast surface motion, which are converted to displacement data using surface volume and optical flow techniques. Diagnostic analysis uses this displacement data to identify underlying breast tissue properties, such as stiffness and damping. Cancerous tissue is 400∼1000% stiffer than healthy tissue and, as such, can provide much higher diagnostic contrast than radio-density used in mammography (5∼10%). Diagnostic algorithms are able to be fully automated, removing requirements for skilled personnel to interpret diagnostic images, providing more consistent diagnosis. Critical analysis of mammography's diagnostic accuracy identified a number of issues with studies including using cohorts inclusive of larger palpable and/or prevalent tumors, resulting in disproportionately high sensitivity. Furthermore, accuracy based on interval cancers presenting between screens is flawed and highly dependent on screening interval resulting in higher than true sensitivity and specificity. The most sound methodology for assessing accuracy came from mammography studies, which compared mammography to other breast imaging modalities, such as ultrasound or MRI, resulting in average sensitivity and specificity values of 60% and 80%, respectively. These values determined the first criteria used to assess the diagnostic performance of DIET diagnostic algorithms, the second was to achieve a highly sensitive diagnostic with sensitivity and specificity 80% and 65%, respectively. A clinical trial involving 14 patients (28 breasts, 13 cancerous, 15 healthy) was carried out using the DIET technology following mammography screening. This clinical cohort is varied breast sizes from 200-1100cm3 and tumor diameters from 7-48 mm, and was used to test di- agnostic algorithms presented in this thesis. This unique clinical dataset is used to test novel diagnostic algorithms presented in this thesis. The core of this thesis presents diagnostic algorithms capable of providing higher diagnostic accuracy than mammography using DIET to realise the many potential benefits of this safe, portable, and non-invasive screening technology. The first diagnostic method uses a model developed in this thesis, based of Rayleigh damping, to describe the viscous damping distribution in the breast. The computationally efficient diagnostic algorithm segments the breast into four radial segments and fits this viscous damping model (VDM) to the viscous damping distribution of reference points in each segment. One model coefficient, related to stiffness, was then compared between segments using percentage thresholds, with healthy breasts hypothesised to have similar coefficient values between segments and cancerous breasts expected to have more varied coefficient values, indicative of a tumor. Twelve breast segmentation configurations were tested to ensure robustness. The optimal configuration, located on the outer more neutral side of the breast, resulted in optimal sensitivity and specificity of 80% and 75%, respectively, with a receiver operator characteristic (ROC) curve area (A-CC) of 0.84. The second diagnostic algorithm involves assessment of response frequency with higher frequency response associated with increased stiffness, and potential tumor presence. A similar segmentation methodology, with increased breast segmentation in the vertical direction was used to attempt to increase diagnostic resolution of smaller tumors. Second dominant response frequencies for all reference points in each segment were averaged, providing frequency component magnitudes were sufficient. Mean frequencies in segments around each vertical band were compared using diagnostic tolerances with segments outside the tolerance indicating high variability in frequency composition and, thus, potential cancer. As with the VDM method, twelve segmentation configurations were tested. Optimal configuration resulted in sensitivity and specificity of 81% and 75%, respectively and ROC curve A-CC of 0.85. Further combining of diagnostic results from two segments on opposite sides of the breast resulted in 100% sensitivity and 69% specificity. All diagnostic criteria were exceeded in both methods and diagnostic accuracy exceeds mammography. Finally, two clinically feasible methods were developed to combine and optimise these individual, tissue mechanics based diagnostic algorithms. The first uses opposite configurations in the frequency decomposition (FD) method. Consistent diagnosis in the two configurations are considered true and inconsistent results are diagnosed using the VDM method. This method results in sensitivity and specificity of 92% and 86%, respectively. The second combined method uses DIET measured breast volume to dictate the method used with small breast volumes diagnosed using the VDM method and large breast volumes diagnosed using the FD method. This method gives 100% sensitivity and 80% specificity. These clinically feasible methods show further diagnostic improvements resulting in a high diagnostic accuracy for this proof-of-concept clinical cohort. These diagnostic algorithms and optimal combinations prove high diagnostic accuracy using DIET can be achieved. This thesis shows implementation of DIET into breast screening programs could provide fast, automated breast cancer diagnosis and consequently faster treatment, improving outcomes and lowering treatment costs. DIET can increase equity for younger women, and positively impact breast screening for M-aori women and women living rurally, with portability enabling mobile screening services to reduce physical barriers to receiving breast care and increase equity of access. These benefits will also impact breast screening programs in developing countries, where mammography is not feasible due to excessive infrastructure cost and requirement for trained radiologists and technicians. Overall, this thesis quantifies the need and potential benefits of a new breast screening solution and critically analyses the diagnostic accuracy presented by mammography, identifying issues of infiated sensitivity values, which are commonly believed to be valid. Most importantly, this thesis develops automated diagnostic methods including combined clinically feasible algorithms, resulting in sensitivity and specificity, which far exceed mammography in this cohort. These diagnostic methods take DIET from a technology with undisputed benefits in terms of screening procedure, to a clinically feasible technology with proven diagnostic potential, thus, worthy of further research and investment.
|All Right Reserved
|Automated breast cancer diagnostics using a Digital Image Elasto Tomography (DIET) system.
|Theses / Dissertations
|University of Canterbury
|Doctor of Philosophy
|Faculty of Engineering
1 - 1 of 1
- Fitzjohn, Jessica_Final PhD Thesis.pdf
- 37.95 MB
- Adobe Portable Document Format
1 - 1 of 1
No Thumbnail Available
- 1.71 KB
- Item-specific license agreed upon to submission