1.

Introduction

Since the first SPIE meeting on medical imaging (MI) in 1971, the conference has played an integral role in the development of countless technologies for MI. That first conference (Quantitative Imagery in the Biomedical Sciences I) covered a range of physics-related topics, including image quality, image perception, and radiography. In the decade after, physics continued to play a central role in SPIE meetings of MI until the word “physics” first appeared in a conference title in 1982 (Physics and Engineering in Medical Imaging). Throughout the 1980s, annual meetings on MI topics continued to include physics through instrumentation, image formation, image processing, image quality, and more. MI became a regular, annual conference in 1987 and quickly expanded to multiple subconferences, where Physics of Medical Imaging became a mainstay since 1993. The conference has been largely informed by proffered papers, reflecting the progression of the field and serving as the premier venue for researchers and scientists to present and discuss important new developments.

The physics topics served MI technology and practice, with progression across modalities, metrologies, technologies, and applications. For example, the early days of linear systems theory described the transfer of imaging information through an analog screen-film x-ray system. These concepts were then applied to the new flat-panel digital detectors that dramatically changed the face of diagnostic radiology in the 1990s. More recently, the introduction of neural networks holds the promise of a similar sea change in the delivery of patient care. SPIE has always been the event where the field has congregated to see the future.

The purpose of this paper is to demonstrate and highlight the major trajectories of imaging physics and how they are informed by the community and science presented (largely but not exclusively) at SPIE MI conferences from its inception to now. In a broad stroke, we identify the field development through SPIE into 11 subfields detailed below (Table 1), which interrelate with each other in many ways. This list is by no means comprehensive, but rather reflective of the progression of the discipline as witnessed and highlighted through SPIE MI. As a result, many categories, modalities, and topics will not be covered due to limited focus and space. In addition, we aim to highlight a representative sample of SPIE MI papers showing the breadth of topics presented at SPIE conferences, but these are only a small sample of the many outstanding papers presented over the years and even then are not chosen to single out the “best” ones. Using the titles of the top five downloaded abstracts from each year, we generated a word cloud that represents the most common words (Fig. 1). As can be seen, these represent many familiar topics including imaging, x-ray, detector, digital, image, radiography, mammography, noise, and more. They highlight many concepts that together form the scaffolding of what we recognize as the physics of medical imaging.

Table 1

Major subfields of imaging physics discussed in this paper.

Fig. 1

Word cloud of most frequent words in SPIE Physics of Medical Imaging paper titles.

2.

Image Science

About 50 years ago, scientists met at the SPIE MI conference to present and discuss new discoveries in image science that represented the start of a knowledge revolution with profound implications for medical care. This exciting field has moved from the maturing of film-based technologies and early discoveries in electronic imaging, through the discovery of physical principles that determine the performance and image quality of new technologies, to recent advances in photon counting technologies, noise reduction methods, and the use of artificial intelligence (AI) for image reconstruction and interpretation. This section highlights some of these key contributions to image science.

In the early 1970s, x-ray image intensifiers (XRII) were used with optical beam splitters to capture moving images both on 35-mm film and analog charge-coupled device (CCD) cameras, and research was starting on quantitative methods to characterize system performance. Early ideas on contrast ratio and modulation transfer function (MTF),¹ focal spot,² and x-ray spectral shape³ were all introduced at SPIE meetings where scientists from the food and drug administration (FDA) were leading the push to develop quantitative metrics for quality assurance in these new technologies. While Fourier ideas were introduced in the 1950s,⁴ practical use of MTF was still a novelty. At the 1979 SPIE meeting, Macovski⁵ introduced the idea of dual-energy imaging and Ovitt et al.⁶ introduced digital subtraction angiography (DSA) including a section on the benefits of using a digital image format. Thus started the digital revolution in image science.

Image science became a serious topic in the 1980s. Particular contributions came from Wagner, Burgess, Shaw, Barrett, Myers, and colleagues including the use of statistical decision theory and signal-to-noise ratio (SNR) concepts. For example, at the 1981 SPIE meeting, Wagner et al.⁷ and Tapiovaara and Wagner⁸ described the importance of the squared-SNR in detection theory and concepts of information transfer efficiency (Fig. 2). These led to widespread use of the detective quantum efficiency (DQE) including the noise-equivalent quanta (NEQ).⁹ At the 1984 SPIE meeting they described a generalized DQE approach that described energy-dependent weights that would improve iodine SNR, something that would not be possible until the development of photon-counting detectors (PCD).⁸ A series of articles and presentations introduced human signal-detection theory to radiographic imaging with the common feature that detection tasks in radiographic images are often noise limited.¹⁰ Much of this work is summarized in an International Commission on Radiation Units & Measurements (ICRU) report including the importance of Fourier methods and connection to detectability index and ideal observers in 1996.¹¹ The idea that quantum-based imaging systems could be represented as a cascade of simple gain and blur processes was introduced.^12,13 Barrett led the development of digital spatial-domain methods to study detectability and system performance that is more comprehensive and appropriate for digital imaging systems.^14–16

Fig. 2

X-ray tube voltage dependence of some radiographic task-dependent DQEs. The curves correspond to 20-cm phantom thickness and are for detecting a small thickness of target material. Reprinted from Ref. 8.

The 1990s and 2000s saw widespread adoption of Fourier and spatial-domain metrics for the description of the image signal and noise,^17,18 including developments in cascaded-systems theory which started in the 1980s to describe the connection between system design and performance metrics.^19,20 The use of Fourier metrics continues to this day, with the modern iterative and nonlinear image processing and noise reduction algorithms necessitating assessment conditions that are needed to satisfy the linearity and shift-invariance requirement of Fourier methods.²¹ New concepts in image science are required to keep pace with the growth of AI to provide the scientific basis of these methods and provide an understanding of their potential to improve detection tasks and optimize image quality.

3.

Image Quality Metrology

Concepts for measuring image quality are intimately tied to developments in image science. At the beginning of the 1970s most of the imaging systems broadly used were still two-dimensional (2D) and based on analog imaging detector configurations, e.g., based on screen-film systems or image intensifier tubes. These were tested regarding their image quality by using Fourier-based metrics such as MTF and DQE, but also low-frequency contrast tests, field evaluation, focus determination, and easy contrast tests as, e.g., presented in 1974.¹ Shaw and van Metter developed a theoretical approach for MTF and DQE measurements of screen-film systems in 1984²² while together with Bunch, they looked into signal-to-noise measurements of such systems,²³ which was in a way following the descriptions by Wagner and Weaver in 1974.²⁴ Wagner and Brown²⁵ presented in 1984 a “more unified analysis of medical system SNR characteristics” (Fig. 3), representing the NEQ of three different screen-film-systems. The approach presented in this paper derived the ideal observer SNR based on statistical decision theory, which linked measurable values to the decision required in radiological imaging. This approach was further followed and investigated.¹⁵ The link between physics-based measurements and the task of detection was established more clearly in the following years.¹⁶ In parallel, studies describing the human observer performance on MI systems were conducted.

Fig. 3

NEQ spectrum for three calcium tungstate screen/film systems at optical density = 1.0. Reprinted from Ref. 25.

Beginning with the broader implementation of CT systems as well as digital imaging detectors for radiography and fluoroscopy, there were new lines of image quality assessment for such systems either based on Fourier-based or spatial-based approaches, both having advantages and disadvantages. It became clear, that the limitations of the digital systems had to be understood and that the image output might depend on the information content to be detected. Thus, papers were looking for the information content of the information patterns.²⁶ To keep pace with the fast development of new types of imaging systems, methods were developed to more practically assess achievable image quality and the ability to provide an estimate of the detectability of lesions or pathology using such imaging systems. An example study evaluated the generalized NEQ and a detectability index for tomosynthesis and cone-beam CT.²⁷ In addition, efforts are being made to measure the image quality or relevant descriptors of it no longer on phantom images or for artificial tasks but in the patient images itself first introduced at SPIE.^28,29 This has also extended to 3D imaging like CT,^30–32 found to be directly relatable to diagnostic performance.³³ It can be hoped that by such approaches, closer links between diagnostic performance and the determined image quality can be gained in the future. This might lead to new concepts for optimization based on targeting acceptable image quality levels.³⁴ This is of great importance given the broad range of developments for new types of imaging systems and image reconstruction methods based on AI. Especially with respect to AI-based methods, new approaches for image quality assessment and quality assurance will be needed.

4.

Radiography

Radiography is the oldest MI modality, and while its earlier developments predate SPIE, the Physics conference has provided many significant contributions. Because of its importance as a modality and relative simplicity in implementation, radiography has long been at the forefront of adopting new technologies and ideas. By 1980, the advantages of digital radiography (DR) were clear, including advantages in dose, scatter reduction, ease of operation, data transmission, image storage, flexible display capability, and image manipulation such as edge enhancement, filtering, and subtraction.³⁵ While analog radiology was still well entrenched and continued to be important,^36,37 DR extended the lessons learned from analog technologies to offer new advantages. The SPIE Conference on Digital Radiography in 1981 already recognized applications in DSA,³⁸ digital tomosynthesis,³⁹ and dual-energy radiography (Fig. 4).⁴⁰ Although the advantages of digital systems were described early on, improving their image quality and their acceptance in clinical practice still took considerable time and effort. For example, pixel size may be an obvious consideration for spatial resolution, but the effects of converter material and thickness, photodiode (PD) gain, and electronic noise all play an important role in image quality. Other considerations include field coverage, uniformity, quantum efficiency, sensitivity, dynamic range, acquisition speed, frame rate, noise characteristics, and cost. Initial DR systems were based on computed radiography (CR), which uses a photostimulable phosphor to capture and store the image, which is then read out with a laser system.^41,42 While CR has a large dynamic range, is digital and is portable, imaging plates must be processed through a reader and intrinsic image quality was seen as lacking.⁴³ Numerous other DR approaches were developed,⁴⁴ but the second generation of DR largely made use of flat-panel detectors (FPD) systems, which directly readout digital images. In addition to the instant readout, the advantages of FPDs in high efficiency and spatial resolution made them almost ideal for radiography.⁴⁵

Fig. 4

Digital pyelographic image (a) and corresponding dual-energy soft tissue subtraction view (b) for a patient with renal carcinoma. Reprinted from Ref. 40.

The developments in radiography were intertwined with developments in image science and detectors (described in those sections). Even today, radiography continues to evolve, with recent advances including dark-field chest radiography⁴⁶ and twin-robotic arms for flexible positioning and setup.⁴⁷

5.

Flat-Panel Detectors

The development of digital FPD (also referred to as flat-panel imagers) has been central to the development of many x-ray based modalities, including radiography, mammography, fluoroscopy, tomosynthesis, and cone-beam CT. The SPIE MI physics conference has always been the platform for presenting original ideas of the design, fabrication, and evaluation of FPD, with over 300 articles on this topic since 1991. With the breadth of knowledge in DR and imaging physics in the first two decades of SPIE MI, it was natural (in 1991) to hear the first presentation on the application of amorphous silicon (a-Si) PD in radiotherapy and diagnostic imaging, which was later referred to as the indirect FPD (Fig. 5).⁴⁸ The following year (1992) the first paper on amorphous selenium (a-Se) based, or direct conversion FPD was presented.⁴⁹ This decade saw a rapid growth of the number of SPIE presentations on both direct and indirect FPD, and an increasing number of contributions from the industry while FPD was being commercialized.^50–55 As FPD technology made its way through clinical translation, image quality evaluation and its standardization became an important topic of the conference.^56,57 Investigations presented at the conference became the foundation for International Electrotechnical Commission and other major standards for the evaluation of FPD.^58,59 Modeling techniques, such as the cascaded linear systems model, for FPD were extensively presented at the conference, and used to optimize FPD designs and identify potential areas of improvement.^60,61 It was realized that the low dose performance of FPD is not x-ray quantum noise limited at the lowest exposures experienced in fluoroscopy. It became a major goal of the community to develop new FPD concepts and materials that could overcome this remaining issue of FPD.

Fig. 5

Schematic representation of a portion of a flat-panel imaging array. (a) Top view showing details of the PD sensors, field-effect transistors (FETs), and data, FET, and bias lines. (b) Cross-sectional view of sensor-FET combination for a single imaging pixel. Reprinted from Ref. 48.

Since 2000, many SPIE MI contributions have focused on novel ideas of how to improve the signal gain to overcome electronic noise in FPD. These ideas fall into two categories: (1) Decrease electronic noise by incorporating pixel amplification, referred to as active pixel sensor (APS); (2) Increase signal gain by using high conversion gain x-ray photoconductors, or optical detectors with avalanche gain. In the former category, pixel amplification was designed incorporating three transistors using either a-Si or polycrystalline Si (Poly-Si)^62,63 with prototype pixels or small arrays fabricated and tested. The lowest noise APS was achieved with crystalline silicon (c-Si) CMOS sensors. This was made possible with the advancement and cost-reduction in wafer-scale CMOS sensors, which were tiled to make FPD for medical x-ray imaging.^64–70

To increase signal gain, avalanche amorphous selenium has been investigated for both direct and indirect detectors.^71,72 For x-ray quantum noise limited performance down to a single x-ray photon, a gain of 10 is sufficient.⁷³ For direct FPD, a variety of direct conversion materials were presented at the SPIE MI conference and drove industrial development for polycrystalline lead oxide (PbO), lead iodide (${\mathrm{PbI}}_2$), mercuric iodide (${\mathrm{HgI}}_2$), and cadmium telluride (CdTe).^74–77 More recently, organic–inorganic perovskite direct conversion FPD began to be presented at SPIE MI,⁷⁸ and we anticipate more work to be presented in future meetings, which will provide guidance for further development necessary for clinical translation and commercialization.

6.

Mammography

The development of mammography over the last 50 years has been truly remarkable. National breast cancer screening programs have been introduced, which has had a huge impact on women’s health, and the technical development has been extensive. In the early 1970s when the first SPIE MI symposium started, national breast cancer screening programs had not started yet; mammographic examinations were performed for diagnostic purposes. The first significant screening program was started in New York in 1963.⁷⁹ A decade later, screening trials were underway or had been carried out in a few different countries, e.g., USA, Sweden, and the UK.^80,81 From that point in time, national screening programs were implemented and became a strong driving force for new inventions in breast imaging. The development of the x-ray detector technology has been an important part of the SPIE MI symposia, together with other inventions in mammography.

X-ray mammography was originally carried out with nonscreen films, to achieve the high spatial resolution required for detecting calcifications. In the 1970s, dedicated screen-film systems for mammography were introduced with a 10 times reduction in dose.^81,82 Early attempts to convert the analog technique to digital included replacing the screen-film system with computed radiography^42,43 and slot-scan detectors.⁸³ Later the detector was replaced by FPDs, with a much faster read-out and significantly better DQE (Fig. 6).^64,84–86 Shortly after the FPDs were introduced, PCDs based on crystalline Si⁸⁷ or multichannel gaseous ionization chamber were invented.⁸⁸ Recently, a new generation of PCD based on GaAs was proposed.⁸⁹

Fig. 6

Clinical mammography images. The magnified inset is an enlargement of the biopsy marker. (a) Indirect conversion CsI + CCD ($40\text{-}\mu \mathrm{m}$ pixels). (b) Direct conversion selenium detector ($70\text{-}\mu \mathrm{m}$ pixels). Reprinted from Ref. 84.

In parallel with the developments of new and improved mammography detectors, other techniques were invented to further increase the diagnostic performance of mammography and in particular screening. In the 1980s, automatic exposure control (AEC) systems and dedicated antiscatter grids for mammography were developed.⁹⁰ The AEC systems consisted originally of an ion chamber that controlled the exposure, whereas, in modern digital systems, the signal from the detector itself controls the exposure. The antiscatter grid prevents to a large extent the scattered radiation from reaching the detector, at the cost of an increased dose to the breast by a factor of 2.5 to 3. Recently, the virtual grids have shown promise to replace physical grids with the potential for reduced dose.⁹¹

As mammography became digital, new inventions became possible. Contrast-enhanced mammography,^92,93 digital breast tomosynthesis,^94–97 and contrast-enhanced breast tomosynthesis,^98,99 as well as breast CT,^100,101 have been invented. All these techniques have shown higher sensitivity and/or specificity than 2D digital mammography. Recently, mechanical imaging has also shown the potential to improve specificity.^102,103

7.

Fluoroscopy

The advent of the XRII in the 1950s marked the adoption of x-ray fluoroscopic imaging and enabled the use of x-ray imaging during interventional and surgical procedures. An XRII consists of a cesium iodide phosphor deposited directly on the photocathode of an intensifier tube. Impinging x-ray photons create electrons that are accelerated and minified until hitting an output phosphor, producing an image $\sim 105$ times brighter than the input image. By capturing the $\sim 2.5\text{-}\mathrm{cm}$ diameter output image with a video camera, an interventionalist could visualize high-contrast devices and contrast-filled vessels in real-time on a TV screen. Some of the earliest work at the SPIE MI conference in the 1970s characterized the image performance of XRIIs.¹ At around the same time, analog-to-digital conversion was applied to the video camera output signal, permitting digital storage and manipulation of x-ray images. This capability led to the first major revolution in fluoroscopy, the invention of DSA.^104,105 Early images published in 1978 are shown in Fig. 7.⁶ Through the 1980s, digital technology also enabled dual-energy subtraction for bone and soft-tissue removal,⁵ and laid the foundation for automated image analysis such as blood-flow (contrast-flow) measurement, device detection, and motion and image-distortion correction. In the same timeframe, further improvements in image quality were achieved by replacing the video camera with solid-state detector technology (e.g., CCD cameras).¹⁰⁶

Fig. 7

(a) Dog’s neck with dilute contrast in the carotid arteries (not visible). (b) Subtracted image showing contrast. (c) Subtracted image with contrast stretching in the display. Reprinted from Ref. 6.

The second major revolution in x-ray fluoroscopic imaging arrived with the replacement of the XRII by the digital FPD. The development of flat panel technology through the 1990s as described above, but it is interesting to note that the SPIE MI conference provided a forum where industry and academia discussed technology developments in an open and cooperative environment that is not often seen in scientific exchanges.⁵³ The transition to a fully digital, geometrically accurate, wide dynamic range, linear signal image acquisition system enabled several important applications for fluoroscopy systems. Quantitation became possible and a significant body of work correcting for detector nonidealities such as lag,¹⁰⁷ bright-burn, nonlinearity, and patient and detector internal scatter was developed. Multimodality integration—that is overlay of 3D information into a real-time 2D fluoroscopic image using rigid or nonrigid registration algorithms—also became a standard component of fluoroscopy-guided minimally invasive procedures, along with automatic device detection, device tracking, and device enhancement algorithms.^108,109 The basics of image quality and dose reduction in x-ray fluoroscopy will always be a focus at SPIE MI conferences; new technologies such as dynamic prepatient collimation, AI-based noise reduction, novel antiscatter grids, and more will continue to enable ever-more-complex minimally invasive image-guided interventions.

8.

Tomosynthesis

About 50 years ago, at the time of the first SPIE MI meeting, tomographic imaging was only available through conventional tomography using film as the image detector. In the early 1970s, a number of investigators started to replace film with image intensifiers thus modifying these analog systems to enable digital capture and analysis. Baily and colleagues¹¹⁰ described a fluoroplanigraphy system in 1971 with the potential for image digitization and subsequent image processing. A decade later, they described the realization of a digitized tomography system.¹¹¹ At the same time, CT scanners in projection (Scout view) mode were used to directly acquire digital radiographs that were used to synthesize arbitrary slice images.³⁹ It was not until the beginning of the new millennium that the development and readiness of flat panel digital detectors capable of rapid image read-out enabled viable tomosynthesis imaging systems to be the subject of widespread research and eventual clinical use. At this time, there was growth in system design and optimization research for musculoskeletal, thoracic, and mammographic applications. In addition, there was significant progress in the development of different reconstruction approaches. These included matrix inversion algorithms,¹¹² filtered back projection,^113,114 and iterative techniques.^94,115

The initial impetus for tomosynthesis imaging was to improve thoracic imaging but this rapidly evolved to include musculoskeletal imaging applications¹¹⁶ and the development of systems that could produce tomosynthesis images of all anatomies.¹¹⁷ Radiographic tomosynthesis research has since progressed to rigorous clinical trials demonstrating efficacy¹¹⁸ and continued efforts to apply new scientific and technical developments to enable novel clinical imaging applications including intraoral dental imaging¹¹⁹ (Fig. 8) and real-time imaging in radiotherapy.¹²⁰

Fig. 8

(a) 2D image of a tooth phantom containing a metallic filling, caries, and fractures. (b) Reconstructed tomosynthesis slice in the plane of the filling and the caries lesion. (c) Reconstructed tomosynthesis slice in the plane of the fractures. Note the increased detail in the lesions (arrows) and alveolar bone in (b) and (c). Reprinted from Ref. 119.

Given the lack of an alternate tomographic imaging modality like CT, it is not surprising that the majority of tomosynthesis research and development has been focused on breast imaging. There continues to be considerable research into system design and optimization including pioneering work on tomosynthesis optimization at SPIE,¹²¹ without a consensus on key factors such as optimal sweep angles, number of projections, and step-and-shoot versus continuous acquisition. These debates are likely to grow with increasing options resulting from the incorporation of new developments in x-ray sources, scanning geometries, reconstruction algorithms, and dose variation and reduction schemes. The need for the standard projection view in the exam and the generation and role of synthesized 2D views from the tomosynthesis data also continue to be investigated. As tomosynthesis has become the standard of care imaging technique for breast cancer screening, research and development efforts are turning from system design and validation to improving image quality, investigating novel imaging geometries and reconstruction approaches, and developing tools and metrics for quality control including the development of dedicated phantoms for evaluating tomosynthesis systems.¹²² Research trends include the expansion of tomosynthesis to include contrast-enhanced exams¹²³ and fusion with other imaging modalities, such as ultrasound.

Distributed x-ray sources provide unique imaging advantages and the opportunity for more flexible system design. Some of the earliest tomosynthesis research and development utilized a scanning-beam x-ray source in an inverse geometry in which the area of the source was much greater than that of the detector.¹²⁴ This system was designed for dynamic, real-time tomographic cardiac imaging. In the last decade, the application of distributed sources to tomosynthesis has become a growing area of research interest.¹²⁵ Undoubtably this trend will continue as distributed source array technology matures and becomes integrated into commercial systems. These source arrays enable greatly increased flexibility in system design and reduced acquisition times—ultimately enabling investigation of new or improved applications in breast and body imaging.

9.

Cone Beam CT

Cone-beam computed tomography (CBCT) is among the important MI technologies for which SPIE MI has been an important forum in reporting early developments and clinical translation. Research at the turn of the century focused on the incorporation of FPDs in novel CBCT scanning systems and set the stage for systems to be developed in the following decades for a spectrum of diagnostic and interventional applications. Early work incorporated an FPD on a CT scanner gantry for possible application in 3D angiography.¹²⁶ CBCT was also developed for image-guided radiation therapy (IGRT),¹²⁷ which grew quickly to represent the standard of care in IGRT. That work was extended to mobile and fixed-room C-arms for CBCT in image-guided surgery and interventional radiology (Fig. 9).^128,129 Applications of CBCT in diagnostic imaging were soon to follow, including CBCT technologies for 3D breast imaging^130,131 as well as musculoskeletal imaging.¹³²

Fig. 9

Various CBCT images: (a) surface rendering of skull phantom, (b) sagittal image of an amputated limb, and (c) transaxial image of a low-contrast phantom with a steel needle inserted in the field of view. Reprinted from Ref. 128.

CBCT image quality was an important topic throughout the development of such emerging systems to identify and mitigate the key factors that limited image quality, such as detector performance, x-ray scatter, and other sources of image noise and artifact. Such work drew from the foundations of MI physics, modeling of the imaging chain, and task-based imaging performance evaluation that were the hallmarks of the Physics of Medical Imaging conference throughout the 1980s (for screen-film systems) and the 1990s (for FPDs). Extending such models to CBCT, fully 3D NPS, NEQ, and task-based detectability index was applied to CBCT systems.¹³³ The role of x-ray scatter as a limiting factor in CBCT was well appreciated, and scatter artifact correction methods ranged from parametric or Monte Carlo modeling¹³⁴ to beam-blocker and related methods.¹³⁵ As CBCT systems found prevalent clinical application in diagnostic and image-guided settings, ongoing work identified patient motion as an important aspect of image quality degradation, with a number of motion compensation methods under development.¹³⁶

As CBCT systems found broad clinical application and their image quality limitations were fully appreciated, novel means of 3D image reconstruction formed an important area of research. Such work proceeded closely in step with similar activity in diagnostic helical CT, as described in greater detail below. Optimization-based alternatives to 3D filtered back projection were an active area of research, including model-based image reconstruction (MBIR) methods with improved noise-resolution tradeoffs. MBIR methods were demonstrated for CBCT,¹³⁷ including predictive models for imaging performance in MBIR¹³⁸ and extension to novel image acquisition protocols—e.g., noncircular orbits.¹³⁹ MBIR models were also extended to incorporate patient-specific prior information (e.g., from a previous CBCT scan) for CBCT applications in image-guided interventions.^140,141 Deep neural networks represent an important emerging basis for CBCT image reconstruction¹⁴² raising important opportunities to improve image quality, to understand the fundamental limits to deep learning performance, and to better inform such approaches by the incorporation of physical models.

10.

Image Reconstruction

Advances in image reconstruction have been critical to the development of diagnostic CT from a 2D parallel-beam imaging modality to the current fully 3D multidetector technology, as well as for the development of CBCT imaging for interventional radiology, oncology, and breast cancer screening and diagnostic. Presentations at the conference have closely followed and enabled these developments. To transition from parallel-beam CT to the current state-of-the-art CT technology, there once was the crucial need to discover how to perform a filtered back-projection reconstruction in the direct fan beam geometry when only a short scan of data is available. The problem was in the partial redundancy of the data, which appeared incompatible with the application of the ramp filter. Giving up on equal data weighting, in favor of utilizing a differentiable data weighting function,¹⁴³ was a major finding that not only supported the targeted application but enabled, through various extensions, the development of single-slice spiral (a.k.a. helical) CT in the early 1990s,^144,145 the expansion to multislice CT at the beginning of the century,¹⁴⁶ and the final transition to the fully 3D geometry of nowadays,^147–149 alongside with a simultaneous expansion of CT applications to include cardiac imaging.^150,151 While judicious data weighting schemes helped produce images of useful quality, it was understood early on that theoretically exact and stable reconstruction would require solving the fully 3D cone-beam image reconstruction problem.^152,153 This problem was extensively investigated for a circular trajectory (Fig. 10),^154–157 and the results have been beneficial for large cone angle diagnostic CT as well as cone beam imaging with FPDs. Down the road, amazing solutions emerged to handle axially truncated data as needed for multislice CT with a helical acquisition,^158,159 which played a major role in terms of algorithm refinements balancing cone-beam artifacts together with efficient use of radiation dose and mitigation of artifacts due to beam hardening and patient motion. Over time, great progress was also made for various other geometries,^160–164 which will support future advances in cone-beam CT with large area detectors.

Fig. 10

Different cone-beam CT reconstructions of simulated phantom. Top row: $xz$ plane, bottom row: $yz$ plane: (a) original phantom, (b) Feldkamp-type half-scan reconstruction, and (c) Grangeat-type half-scan reconstruction. Reprinted from Ref. 154.

Many applications of x-ray CT occur in a busy clinical environment, where fast workflow is a necessity. For this reason, image reconstruction scientists have also devoted important effort to the development of fast algorithms, by using more efficient mathematical reformulations,^165–168 or by identifying ways to benefit from novel computing hardware.^169–173 Such developments have been and remain critical as CT image reconstruction is evolving from linear to nonlinear methods, either with optimization techniques or deep learning paradigms or with a combination of these tools.^174–178 While optimization methods are still under investigation for CT, it is worth mentioning that they have been used for more than two decades in PET, due to the much smaller dataset size, with important advances then presented at the conference.^179–181 Last, we note that the conference served as an important venue to support image reconstruction developments in tomosynthesis for breast imaging, with contributions spanning dedicated analytical formulas,¹¹⁴ optimization techniques,¹⁸² and deep learning approaches.¹⁸³

11.

Diagnostic CT

Since its introduction in 1972, CT has advanced at a rapid pace and is now a pillar of radiological diagnosis. The first CT systems were mere head scanners, with an acquisition time of several minutes per image. In the 1980s whole-body CT scanners became available, with scan times per image reduced to a couple of seconds and significantly improved spatial resolution. However, only single axial slices of the patients’ anatomy could be acquired. The ability of continuous rotation led to the development of spiral (helical) CT in the early nineties (Fig. 11)—this was a significant breakthrough in the history of CT.^144,184,185 For the first time, volume data became available without the risk of mis- or double-registration of anatomical details, albeit initially with limited scan ranges. Larger volume coverage in shorter scan times with thinner image slices was made possible by the widespread introduction of multidetector row computed tomography (MDCT) by all major CT manufacturers in 1998.^186,187 As a result of faster gantry rotation, MDCT also expanded into areas previously considered outside the scope of third-generation CT scanners, such as ECG-controlled cardiac imaging.¹⁸⁸ Interestingly, the basis for CT image reconstruction with optimized temporal resolution was already laid in the early 1980s.¹⁴³ The first decade of the 20th century was characterized by a race for more and more detector rows—4, 16, 32, 64-row scanners became available at a rapid pace. Consequently, new image reconstruction techniques came into focus that could cope with the increasing cone angles of the measurement rays.^189,190 Iterative reconstruction was investigated as a method to better account for the statistical properties of the CT data.¹⁹¹

Fig. 11

Schematic diagram for constant speed helical scanning (CSH), where $t$ indicates time. The gantry rotational angle $\beta$ is shown in (a) and the state of the x-ray source is shown in (b). The table location $z$ is shown for CSH in (c). Reprinted from Ref. 144.

Increasing clinical experience with MDCT indicated that adding even more detector rows alone would not result in greater clinical benefit. Instead, new CT concepts were evaluated to overcome the remaining limitations of MDCT. While several innovative CT designs never found their way into clinical routine,¹⁹² two concepts held their own over the years: wide-area detector CT and dual-source CT. As a method of extending CT from purely morphologic to functional imaging, spectral CT imaging has experienced a renaissance. New reconstruction schemes¹⁹³ were proposed, techniques to acquire dual-energy CT data were evaluated,¹⁹⁴ and different ways to produce meaningful clinical results were investigated.^195–198 Most recently, another novel CT concept has gained increasing interest in the CT research community: photon-counting CT (described in greater detail below). The first results of photon counting CT systems with increasing maturity demonstrated potential clinical benefits (Fig. 12).^199–202 As the latest exciting development, machine learning has now also made its way into CT image reconstruction and is currently drawing a great deal of attention.^203,204

Fig. 12

Chest patient images from conventional energy integrating detector (EID) CT and PCD CT comparing pathological features such as cystic air spaces (red arrows, upper row) and lung nodules (red arrows, bottom row). The conspicuity of anatomic details noticeably improved in PCD-CT images. The PCD-CT image also exhibited lower image noise than EID-CT at matched acquisition dose as shown in the ROI measurements. Reprinted from Ref. 199.

12.

Photon Counting Detectors

Photon counting is an intuitive way of detecting x-rays, counting them one by one, and measuring the energy for each x-ray using programmable thresholds. In nuclear imaging, such as single-photon emission computed tomography and PET, this has been the gold standard from the start. One advantage compared with energy integrating x-ray detectors is that electronic noise can be eliminated, enabling lower-dose imaging. The second advantage is that since the energy is measured for each x-ray, the information can be weighted to optimize the image quality when the image is formed. The energy measurement can also be used for improved separation between calcium and iodine, for example. Today’s dual-energy solutions have significantly worse energy resolution and/or require unfavorable trade-offs in spatial resolution and workflow. PCDs are direct conversion with reduced pixel size, drastically increasing the spatial resolution. The first paper describing a PCD⁸⁷ that led to an FDA-cleared x-ray imaging system was presented at SPIE MI 2000 (Fig. 13). At that time, it was the only photon-counting paper at the conference. Since then, photon counting has grown into a major topic at the MI conference, and a large part of the progress in this field of research can be attributed to presentations and discussions in that forum.

Fig. 13

Dose-efficient x-ray system with edge-on silicon PCDs for digital mammography. Reprinted from Ref. 87.

One important area of progress has been on the hardware side with improved sensors and fast, low-noise application-specific integrated circuits.^{201,205–207}

A major challenge with photon counting is the high incident rate of x-rays, and the following pile-up of events and charge sharing between pixels. Several contributions helped both to evaluate the impact on image quality and to suggest correction algorithms or detector design strategies to mitigate the problems.^208–212 Other noteworthy contributions investigated the potential for photon counting in terms of energy weighting and material basis decomposition.^213–216 Several promising clinical applications have been suggested, such as dual contrast imaging of the liver, kidney stones, and crystal-related arthropathies, cardiovascular imaging, thoracic imaging, and neuroradiology.^{199,217–220} Two major technology platforms are currently being developed for photon-counting CT.²²¹ The first uses a cadmium-based (CdTe or CZT) sensor with a thickness of a few millimeters.¹⁹⁹ The second is based on so-called “deep silicon,” with sensors being mounted in an edge-on geometry with several centimeters of depth. One system recently received FDA clearance, and we can expect regulatory clearance for several more systems in the next few years. Photon counting also paves the way for future clinical applications such as phase-contrast imaging. Ultrahigh spatial resolution down to $1\mu \mathrm{m}$ can be feasible,²²² and a pseudo-monochromatic beam can be achieved since the x-ray wavelength is measured.²²³

13.

Conclusions

As these MI technologies developed from primitive concepts and prototypes to competing for commercial systems in routine use for many clinical applications, the SPIE MI meeting has served as an effective interface between the basic research of academic scientists, technology development of engineers, and the dissemination of industry research and development. The SPIE MI community played an instrumental role in many key developments in the past 50 years by sharing new ideas and results, learning from each other—ranging from experts in short courses to students giving talks and presenting posters, and forming lasting friendships and collaborations. That journey has never been simple. For today’s community, it may be difficult to appreciate that the very concept of a “digital” MTF or DQE was controversial in the mid 1990s, yet today it is acknowledged as the methodology that accurately describes digital system performance and allows for system performance optimization. The recent discussions on neural networks, their configuration, the issues associated with their “learning/training,” and the evaluation of their performance bear the same hallmarks of the birth of another game-changing technology. The SPIE MI conference will undoubtedly play a pivotal role in the development of this nascent field and many others that hold so much promise for the future betterment of patient care, and we are excited about what the next 50 years of SPIE MI will bring.