Breast cancer is leading cancer among females across the globe. Currently available imaging modalities suffer from low specificity and sensitivity especially for women with radiographically dense breasts. In recent years, the diffuse optical tomography (DOT) that utilized NIR light has emerged as a promising modality for breast cancer screening and diagnosis. Yet, the DOT is not fully exploiting its clinical utility because of unavailability of the reliable and efficient imaging system and also the relatively slow and imprecise image reconstruction process.
In this work, firstly we developed a state-of-the-art clinical diffuse optical tomography (DOT) imaging system which collects multichannel, multispectral, frequency-domain breast data under mammographically compressed geometry. The instrument utilizes APD detectors under homodyne detection scheme to measure the amplitude and phase of the diffuse photon density wave (DPDW). The novel features of the DOT system are (1) concurrent data acquisition under same geometry with digital breast tomosynthesis (DBT), which can be used for volume segmentation for DOT image reconstruction without doing image registration, (2) design of source and detection paddle which permits us to design patient-specific source-detector arrangements, (3) and rapid data acquisition. We validated the DOT system using a number of tissue simulating phantoms.
Secondly, we presented an innovative methodology for optimization of sampling schemes in diffuse optical tomography (DOT). The proposed method exploits singular value decomposition (SVD) of the sensitivity matrix, or weight matrix, in DOT. Two mathematical metrics are introduced to assess and determine the optimum source-detector measurement configuration in terms of data correlation and image space resolution. The key idea of the work is to weight each data measurement or rows in the sensitivity matrix, and similarly to weight each unknown image basis, or columns in the sensitivity matrix, according to their contribution to the rank of the sensitivity matrix, respectively. The proposed metrics offer a new perspective on the data sampling and provide an efficient way of optimizing the sampling schemes in DOT. We evaluated various acquisition geometries often used in DOT by use of the proposed metrics. By iteratively selecting an optimal sparse set of data measurements, we showed that one can design a DOT scanning protocol that provides essentially the same image quality at much-reduced sampling.
Thirdly, we presented a novel convolutional neural network (CNN) based approach to learn the bulk optical parameters of a highly scattering medium for instance biological tissue from the DOT boundary measurements. We validated the proposed method against simulation as well as experimental data. We compared the results of the proposed method with existing approaches for evaluating its effectiveness. The encouraging results demonstrate that our proposed CNN based approach for bulk optical property estimation is not only taken lesser time but also outperform the existing methods in terms of accuracy.
Finally, we presented a new data specific mask guided reconstruction algorithm for diffuse optical tomography (DOT). To address the ill-posedness of the inverse problem, conventional image reconstruction approaches employed regularization with a constant penalty parameter, which uniformly smoothes out the solution. We proposed two methods in this work. The first method exploits a data specific prior mask to design a spatially varying regularization. While the second method imposes a region of interest constraints using the prior mask. Moreover, the algorithm in the second method iterates between the discrete and continuous step to update the mask and the optical parameters, respectively. The prior mask is created from the DOT data itself by exploiting the multi-measurement vector formulation. The results of the study indicate the enhancement in accuracy of optical contrast, better spatial resolution and reduction in noise compared with conventional method.