Browsing by Author "Lo, Joseph Yuan-Chieh"

Breast cancer screening is currently performed by mammography, which is limited by overlying anatomy and dense breast tissue. Computer aided detection (CADe) systems can serve as a double reader to improve radiologist performance. Tomosynthesis is a limited-angle cone-beam x-ray imaging modality that is currently being investigated to overcome mammography's limitations. CADe systems will play a crucial role to enhance workflow and performance for breast tomosynthesis.

The purpose of this work was to develop unique CADe algorithms for breast tomosynthesis reconstructed volumes. Unlike traditional CADe algorithms which rely on segmentation followed by feature extraction, selection and merging, this dissertation instead adopts information theory principles which are more robust. Information theory relies entirely on the statistical properties of an image and makes no assumptions about underlying distributions and is thus advantageous for smaller datasets such those currently used for all tomosynthesis CADe studies.

The proposed algorithm has two 2 stages (1) initial candidate generation of suspicious locations (2) false positive reduction. Images were accrued from 250 human subjects. In the first stage, initial suspicious locations were first isolated in the 25 projection images per subject acquired by the tomosynthesis system. Only these suspicious locations were reconstructed to yield 3D Volumes of Interest (VOI). For the second stage of the algorithm false positive reduction was then done in three ways: (1) using only the central slice of the VOI containing the largest cross-section of the mass, (2) using the entire volume, and (3) making decisions on a per slice basis and then combining those decisions using either a linear discriminant or decision fusion. A 92% sensitivity was achieved by all three approaches with 4.4 FPs / volume for approach 1, 3.9 for the second approach and 2.5 for the slice-by-slice based algorithm using decision fusion.

We have therefore developed a novel CADe algorithm for breast tomosynthesis. The techniques uses an information theory approach to achieve very high sensitivity for cancer detection while effectively minimizing false positives.

This work seeks to utilize a cohort of computational, patient-based breast phantoms and anthropomorphic lesions inserted therein to determine trends in breast lesion detectability as a function of several clincally relevant variables. One of the measures of local density proposed gives rise to a statisically signicant trend in lesion detectability, and it is apparent that lesion type is also a predictor of relative detectability.

Knowledge based radiation therapy planning was developed as a way to use a knowledge base of clinically approved radiation therapy plans to shorten the amount of time necessary to create a new treatment plan. The method has been tested using plans from a database of approved plans from Duke University and it was shown that the quality of the new plan is at least comparable to the original plan. When a match is found from the database for a query anatomy, the fluences of the match plan are deformed to fit the query planning target volume (PTV) and brought into the new query plan. The deformation is currently done using several two dimensional deformation registrations of the beam's eye vies (BEV) projection of the match PTV to the BEV of the query PTV for each beam in the plan. We investigated whether using information from a three dimensional deformation of the PTVs to deform the fluences would yield higher quality plans than the two dimensional method. Using Velocity AI's three dimensional deformation capabilities we deformed several match query pairs and projected the deformation field down into two dimensions for each beam angle. Using the projected fields we deformed each beam's fluence and imported the new fluences into a radiation therapy plan. After evaluating the DVHs of both pre- and post-optimized plans we concluded that there is no benefit to performing the three dimensional deformation method over the two dimensional method for prostate patients.

Intensity-modulated radiotherapy (IMRT) remains the standard of care for external beam radiation therapy for head and neck cancers. Planning for IMRT requires a trial-and-error approach that is completely dependent on planner expertise and time available for multiple iterations of manual optimization adjustments. Knowledge-based radiation therapy planning utilizes a database of previously planned Duke University Medical Center patient plans to create clinically comparable treatment plans by comparing the geometrical two-dimensional projections of the planning target volume (PTV) and organs at risk (OAR). These 2D beam's eye view (BEV) images are first aligned with squared error registration, then the similarity is computed using the mutual information (MI) metric. After the closest match is found, computed constraints and deformed fluence maps are entered into Eclipse treatment planning system to generate the new knowledge-based treatment plan. For this study, 20 randomly selected cases were matched against a database of 103 head and neck cancer cases. The resulting new plans were compared to their clinically planned counterparts. For these 20 cases, 13 proved to be dosimetrically comparable by evaluation of the PTV dose-volume histogram. In 92% of cases planned, at least half of the OARs were also deemed comparable or better than the original plan. These cases were planned in less than 25 minutes with no manual constraint objective adjustments, as opposed to many hours needed in clinical planning.

Intensity-modulated radiotherapy (IMRT) has become an effective tool for cancer treatment with radiation. However, even expert radiation planners still need to spend a substantial amount of time, approximately 4 hours, manually adjusting IMRT optimization parameters such as dose limits and costlet weights in order to obtain a clinically acceptable plan. Also, the quality of the treatment plan generated is solely based on the experience and training of the planning. In comparing the geometries of the planning target volume (PTV), bladder, rectum, right and left femoral heads, a knowledge-based approach to IMRT treatment planning may reduce the time needed to generate a clinically acceptable prostate plan. The knowledge-based approach uses the clinically acceptable plans of previously irradiated patients which are adapted to the new patient. Patient selection is done by using mutual information (MI). Having selected the best matched patient, Elastix (a toolkit for rigid and deformable registration) is used to deform the treatment plan of the previously irradiated patient to the new patient's geometry. The Eclipse treatment planning system is used to generate both pre-optimized and post optimized plans for the new patients. The knowledge-based treatment plans require no manual intervention. For the 101 patient data, it was shown that the newly generated plans were of similar or slightly worse dosimetric quality and were only generated in less than 30 minutes. Given the large size of this data set, the results are likely to be robust in representing treatment planning efficacy over a diverse range of patient anatomy. The results also show that this work has the potential to automatically provide high quality treatment plans while dramatically reducing the dependence of the expertise of the planner and the treatment planning time.

Neutron stimulated emission computed tomography (NSECT) is a non-invasive, tomographic imaging technique with the ability to locate and quantify elemental concentration in a tissue sample. Previous studies have shown that NSECT has the ability to differentiate between benign and malignant tissue and diagnose liver iron overload while using a neutron beam tomographic acquisition protocol followed by iterative image reconstruction. These studies have shown that moderate concentrations of iron can be detected in the liver with moderate dose levels and long scan times. However, a low-dose, reduced scan time technique to differentiate various liver diseases has not been tested. As with other imaging modalities, the performance of NSECT in detecting different diseases while reducing dose and scan time will depend on the acquisition techniques and parameters that are used to scan the patients. In order to optimize a clinical liver imaging system based on NSECT, it is important to implement low-dose techniques and evaluate their feasibility, sensitivity, specificity and accuracy by analyzing the generated liver images from a patient population. This research work proposes to use Monte-Carlo simulations to optimize a clinical NSECT system for detection, localization, quantification and classification of liver diseases. This project has been divided into three parts; (a) implement two novel acquisition techniques for dose reduction, (b) modify MLEM iterative image reconstruction algorithm to incorporate the new acquisition techniques and (c) evaluate the performance of this combined technique on a simulated patient population.

The two dose-reduction, acquisition techniques that have been implemented are; (i) use of a single angle scanning, multi-detector acquisition system and (ii) the neutron-time resolved imaging (n-TRI) technique. In n-TRI, the NSECT signal has been resolved in time by a function of the speed of the incident neutron beam and this information has been used to locate the liver lesions in the tissue. These changes in the acquisition system have been incorporated and used to modify MLEM iterative image reconstruction algorithm to generate liver images. The liver images are generated from sinograms acquired by the simulated n-TRI based NSECT scanner from a simulated patient population.

The simulated patient population has patients of different sizes, with different liver diseases, multiple lesions with different sizes and locations in the liver. The NSECT images generated from this population have been used to validate the liver imaging system developed in this project. Statistical tests such as ROC and student t-tests have been used to evaluate this system. The overall improvement in dose and scan time as compared to the NSECT tomographic system have been calculated to verify the improvement in the imaging system. The patient dose was calculated by measuring the energy deposited by the neutron beam in the liver and surrounding body tissue. The scan time was calculated by measuring the time required by a neutron source to produce the neutron fluence required to generate a clinically viable NSECT image.

Simulation studies indicate that this NSECT system can detect, locate, quantify and classify liver lesions in different sized patients. The n-TRI imaging technique can detect lesions with wet iron concentration of 0.5 mg/g or higher in liver tissue in patients with 30 cm torso and can quantify lesions at 0.3 ns timing resolution with errors ≤ 17.8%. The NSECT system can localize and classify liver lesions of hemochromatosis, hepatocellular carcinoma, fatty liver tissue and cirrhotic liver tissue based on bulk and trace element concentrations. In a small patient with a torso major axis of 30 cm, the n-TRI based liver imaging technique can localize 91.67% of all lesions and classify lesions with an accuracy of 88.23%. The dose to the small patient is 0.37 mSv a reduction of 39.9% as compared to the NSECT tomographic system and scan times are comparable to that of an abdominal MRI scan. In a bigger patient with a torso major axis of 50cm, the n-TRI based technique can detect 75% of the lesions, while localizing 66.67% of the lesions, the accuracy of classification is 76.47%. The effective dose equivalent delivered to the larger patient is 1.57 mSv for a 68.8% decrease in dose as compared to a tomographic NSECT system.

The research performed for this dissertation has two important outcomes. First, it demonstrates that NSECT has the clinical potential for detection, localization and classification of liver diseases in patients. Second, it provides a validation of the simulation of a novel low-dose liver imaging technique which can be used to guide future development and experimental implementation of the technique.

Currently the most common and effective procedure of early detection of breast cancer is through screening mammography. Mammography detects not only invasive cancers, but also in situ lesions including ductal carcinoma in situ (DCIS), which is the most common non-invasive breast cancer. The focus of this dissertation is to investigate machine learning and deep learning technologies to improve detection and diagnosis of breast cancer from mammography, focusing on calcifications that represent a common type of suspicious lesions. This dissertation is composed of four main projects.

The first project developed a transfer learning framework to predict occult invasive disease in DCIS by a machine learning approach. Performance was improved by using forced labeling and domain adaptation based upon data from two related diseases, atypical ductal hyperplasia (ADH) and invasive ductal carcinoma (IDC).

The second project expanded the dataset fivefold, producing the largest cohort of DCIS cases to date. This rich dataset enabled rigorous studies to assess the performance and clinical utility for guiding patients with DCIS who are considering active surveillance or surgical planning.

The third project developed a one-class anomaly detection system to detect calcifications with deep autoencoder network and dissimilarity index. Calcifications were detected by learning negative-only images. Using a very large mammography dataset, this study was one of the first in medical imaging to show model capacity changes based on different architectures as well as different number or type of training images.

Finally, the fourth project broadened the diagnostic task to classify breast calcifications as benign or malignant using a deep multitask network. Unlike previous projects, algorithms here were designed to detect calcifications and classify cancers simultaneously. The multitask network model’s results showed improved performance compared to single classification networks. A correlation between two output losses’ weighting factors was also achieved by analyzing performance impacts from both weighting factors.

In conclusion, this dissertation provides three main contributions. First, it provides a complete classification framework to distinguish invasive disease from DCIS based on the largest single institution dataset. Second, it demonstrates the potential of an anomaly detection approach to detect mammographic calcifications that may facilitate new clinical applications such as computer-aided triage and quality improvement. Third, a unified approach for multitask learning was used to perform detection and classification simultaneously, which provided the highest performance for classification from our group to date.

Breast cancer is currently the most common non-skin cancer and the second leading cause of cancer-related death in women here in the United States. X-ray mammography is currently the standard clinical imaging modality for breast cancer screening and diagnosis due to its high sensitivity and resolution at a low patient dose. With the advancement of breast imaging from analog to digital, quantitative measurements rather than qualitative assessments can be made from these images. One such measurement, mammographic breast density (i.e. the percentage of the entire breast volume that is taken up by dense glandular tissue), has been shown to be a biomarker well correlated with cancer risk. However, a digital mammogram still suffers from its projective nature. The resulting overlap of normal breast tissue can obscure lesions, limit quantitative measurement accuracy, and present false alarms leading to unnecessary recall studies. To address this key limitation, several 3D imaging techniques have been developed such as breast magnetic resonance imaging (MRI), dedicated breast computed tomography (CT), and digital breast tomosynthesis (tomo). Perhaps the most recently developed modality is tomo, which is a limited-angle cone-beam CT of the breast compressed in the same geometry as mammography. Because tomo retains all the aforementioned advantages of mammography but adds depth information and can be built based on an existing digital mammography device, measuring breast density and extracting other quantitative features from tomo images was a major focus of this study.

Before attempting to measure breast density and other features from reconstructed tomo image volumes, the quantitative potential of this imaging modality was assessed. First, we explored a slice-by-slice technique that measures tissue density using only the information from a single slice from the reconstructed tomosynthesis volume with geometrically simple tissue-equivalent phantoms. Once this task has been satisfactorily performed, we studied a probabilistic approach toward quantitation of the entire 3D volume. Some work has been done previously in the realm of 2D hidden Markov random fields (HMRFs) to categorize mammograms according to their Wolf pattern, detect mammographic lesions, and segment satellite and mixed media (text/photograph) images. For this project, a 3D hidden Markov model (HMM) method was developed and applied to tomo images under the simplified assumption that the possible tissue type of each tomo voxel is either adipose (fatty) or glandular (dense). Because adipose and glandular tissue is easily distinguished in MR images, patient breast MRIs were used to train, validate, and finally to assess the accuracy of our HMM segmentation algorithm when applied to tomo images by comparing the volumetric breast density to the MRI breast density for the same patient. Because they are so often studied conjunctively, several image texture features were calculated and compared between MRI and tomo as well.

Another aim of our study was to investigate whether changes in macroscopic 3D imaging features (texture and density) can accurately predict the chemoprevention response that was measured with Random Periareolar Fine Needle Aspiration (RPFNA) cytology for a uniquely young high-risk cohort of women. This aim to investigate the potential of combining multi-modality macroscopic 3D imaging information with a cytological measure of risk and then investigating how response to tamoxifen and other chemoprevention treatment affects each of these risk biomarkers in young, high-risk women is completely novel in the fields of medical physics and biomedical engineering.

Quantitative imaging is important in medical imaging. Physical phantoms are used. There is reason to believe that anthropomorphic physical phatoms are better than uniform phantoms. To investigate this question, we develop a novel imaging metrology with a phatient-based phantom and apply its use to several digital breast tomosytneshis machines. At the same time, we use the traditional means of assessing image quality. Our results show a strong dependence on image performance with the type of phantom used. Furthermore, we demonstrate the feasibility of this metrology in real, clinical applications.

Patients undergoing cardiac catheterization are potentially at risk of radiation-induced health effects from the interventional fluoroscopic X-ray imaging used throughout the clinical procedure. The amount of radiation exposure is highly dependent on the complexity of the procedure and the level of optimization in imaging parameters applied by the clinician. For cardiac catheterization, patient radiation dosimetry, for key organs as well as whole-body effective, is challenging due to the lack of fixed imaging protocols, unlike other common X-ray based imaging modalities.

Pediatric patients are at a greater risk compared to adults due to their greater cellular radio-sensitivities as well as longer remaining life-expectancy following the radiation exposure. In terms of radiation dosimetry, they are often more challenging due to greater variation in body size, which often triggers a wider range of imaging parameters in modern imaging systems with automatic dose rate modulation.

The overall objective of this dissertation was to develop a comprehensive method of radiation dose estimation for pediatric patients undergoing cardiac catheterization. In this dissertation, the research is divided into two main parts: the Physics Component and the Clinical Component. A proof-of-principle study focused on two patient age groups (Newborn and Five-year-old), one popular biplane imaging system, and the clinical practice of two pediatric cardiologists at one large academic medical center.

The Physics Component includes experiments relevant to the physical measurement of patient organ dose using high-sensitivity MOSFET dosimeters placed in anthropomorphic pediatric phantoms.

First, the three-dimensional angular dependence of MOSFET detectors in scatter medium under fluoroscopic irradiation was characterized. A custom-made spherical scatter phantom was used to measure response variations in three-dimensional angular orientations. The results were to be used as angular dependence correction factors for the MOSFET organ dose measurements in the following studies. Minor angular dependence (< ±20% at all angles tested, < ±10% at clinically relevant angles in cardiac catheterization) was observed.

Second, the cardiac dose for common fluoroscopic imaging techniques for pediatric patients in the two age groups was measured. Imaging technique settings with variations of individual key imaging parameters were tested to observe the quantitative effect of imaging optimization or lack thereof. Along with each measurement, the two standard system output indices, the Air Kerma (AK) and Dose-Area Product (DAP), were also recorded and compared to the measured cardiac and skin doses – the lack of correlation between the indices and the organ doses shed light to the substantial limitation of the indices in representing patient radiation dose, at least within the scope of this dissertation.

Third, the effective dose (ED) for Posterior-Anterior and Lateral fluoroscopic imaging techniques for pediatric patients in the two age groups was determined. In addition, the dosimetric effect of removing the anti-scatter grid was studied, for which a factor-of-two ED rate reduction was observed for the imaging techniques.

The Clinical Component involved analytical research to develop a validated retrospective cardiac dose reconstruction formulation and to propose the new Optimization Index which evaluates the level of optimization of the clinician’s imaging usage during a procedure; and small sample group of actual procedures were used to demonstrate applicability of these formulations.

In its entirety, the research represents a first-of-its-kind comprehensive approach in radiation dosimetry for pediatric cardiac catheterization; and separately, it is also modular enough that each individual section can serve as study templates for small-scale dosimetric studies of similar purposes. The data collected and algorithmic formulations developed can be of use in areas of personalized patient dosimetry, clinician training, image quality studies and radiation-associated health effect research.

Breast cancer is the second leading cause of cancer death among women, with approximately 43,780 related deaths annually. Effective screening programs are essential in reducing mortality rates by providing early diagnosis and treatment. Historically, mammography has been the most reliable screening method, significantly dropping the per capita mortality rate since its widespread adoption in the 1980s. However, the increasing workload of approximately 40 million mammography procedures conducted annually in the US poses a significant challenge to the healthcare system. This leads to reports of high rates of burnout among breast radiologists, which can decrease reading accuracy and patient care quality. Therefore, there is a need to improve the existing breast cancer screening workflow to address this challenge.

CAD algorithms have been developed to reduce radiologists' workload and address burnout. However, existing single-view CAD systems often have limited cancer detection performance as well as clinical impact. To overcome this limitation, we collaborated with iCAD (Nashua, NA) to develop a novel Computer-Aided-Detection (CAD) framework for digital breast tomosynthesis (DBT) that mimics the multi-view mammography reading practice used by breast radiologists. As of May 2023, the algorithm is under the initial submission for FDA 510(k) approval. This dissertation introduces a multi-view DBT lesion detection framework consisting of four chapters. Chapter 1 highlights the challenges of breast cancer screening and the necessity for an enhanced DBT CAD algorithm. Chapter 2 presents the single-view detection pipeline, while chapter 3 proposes the ipsilateral refinement concept that improves cancer lesion detection performance. Chapter 4 outlines the temporal matching concept that enhances system-level performance by integrating lesion temporal growth information. Chapter 5 showcases a few additional studies that supported the development of the multi-view lesion detection algorithm.

Our design uses cascaded task-specific models for each of our proposed modules, enabling intermediate reasoning of the multi-view reading steps. This approach allows the radiologist to inspect the output generated by the CAD system and verify the reasoning behind the system's decision, providing an additional layer of validation for complex high-stakes decisions.

Physical phantoms are an important tool in clinical system evaluation. There exists a lack of suitable anthropomorphic physical phantoms that vary as much as a typical patient population. The lack in diversity in anthropomorphic physical phantoms makes generalizing results found using these phantoms difficult. In order to address this issue, a diverse selection of breast phantoms were 3D printed on a Stratasys Objet350 Connex printer using tissue-approximate photopolymers. These cases were then evaluated on a clinical Hologic Selenia Dimensions Digital Breast Tomosynthesis system. The evaluation consisted of a 4-alternative-forced-choice task printed on a contrast insert with silver-doped ink of concentration 200 mg/mL. The disks ranged in size from 350um-770um and a range of signal intensities was achieved by repeatedly overprinting, layering the ink. Each ink pass corresponded to an increase in signal of 1.4%. The contrast insert was imaged in 8 different orientations, at a fixed kVp of 36, and varied mAs for indicated AGD of 1.4, 2.8, and 4.2 mGy. A channelized-Hotelling observer with Gabor channels was used for evaluation and a percent correct was determined. Detection performance increased as dose increased for all cases. The most dense breast case had the worst detection performance as is had the most overlapping structures to obscure the signal. The approximately average density breast and the fatty, thinner breast performed similarly, however this may be due to the beam filtering used to avoid overexposing the detector with the high kVp and mAs used for this experiment. These results indicate that system performance is dependent on the anatomy being imaged. Further investigations with more phantom cases is needed to better evaluate the anatomical dependence of the system performance.