A large area CMOS Active Pixel Sensor for imaging in proton therapy

There is an increased interest in developing imaging systems in proton therapy, with the aim of reducing range uncertainty in treatment planning, assisting in patient positioning and verifying anatomical changes at the time of the treatment. Recently, the PRaVDA collaboration has developed two different solid-state detector technologies for imaging in proton therapy: Silicon Strip Detectors and CMOS Active Pixel Sensors (APSs). This paper reports on the design and optimisation process of the PRaVDA CMOS APSs. Optimisation of parameters, such as epitaxial thickness and resistivity, and performance for individual proton detection and proton radiography are reported.


Introduction
The potential benefits of proton Computed Tomography (pCT) and proton radiography (pRG) to reduce range uncertainty in treatment planning, as well as assisting patient positioning and verifying anatomical changes at the time of treatment [1,2], have driven a world-wide interest in the development of instrumentation for imaging in proton therapy. Several pCT and pRG prototypes have been developed over the last decade, based on a number of different technologies, including Silicon Strip Detectors (SSDs) [3][4][5], scintillating fibres [6][7][8][9], scintillator-based calorimeters [3][4][5][6][7][8]10] and CMOS Active Pixel Sensors (APSs) [11][12][13]. In this scenario, the PRaVDA collaboration was formed to develop a novel instrument for proton imaging and associated reconstruction algorithms. The PRaVDA pCT system [14], currently entirely based on Silicon Strip Detectors (SSDs), features two trackers (four sets of x-u-v SSDs) to measure proton trajectories before and after the patient [15,16], and a mono-dimensional SSD-based energy-Range Telescope (RT) working as a range counter, to infer the residual proton energy from their range.
Although the final choice for the PRaVDA prototype was SSDs, two different solid-state technologies were developed during the course of this project to be potentially used as a range detector: SSDs and CMOS APSs. APSs and SSDs, as position sensitive detectors, allow for many protons to be reconstructed per frame time, unlike calorimeters or scintillator stack designs, which require only one proton per scintillator element during a readout cycle. Both these technologies offer a number of pros and cons when used as an energy detector for pCT.
-1 -While SSDs, measuring 1-D coordinates in each layer, need reconstruction across multiple layers to obtain a 2-D hit position [15], CMOS imagers, being pixelated, have the advantage of providing 2-D coordinates at each detection plane. CMOS imagers, benefiting from advances in lithographic processes and high levels of integration [17], can now be built over large areas (12-inches wafers) by reticule stitching [18,19]. Larger imaging areas needed in pCT and pRG (i.e. 30×30 cm 2 for a head pCT scan) can be easily achieved by mosaic tiling of edge-less CMOS sensors, unlike SSDs. On the other side, SSDs can be designed to sustain radiation damage in very demanding environments (e.g. high luminosity colliders), while CMOS imagers, with appropriate design choices [20][21][22], can offer a moderate radiation hardness in the clinical environment. Different readout speeds can also be envisaged for these two technologies. In fact, SSD can be readout at MHz frequency while readout speed for large area CMOS imagers tends to be orders of magnitude lower. A low readout speed for CMOS sensors, together with the need to keep pCT scan time at clinically feasible lengths, makes the need for track reconstruction across the RT at high occupancy.
Previous works within the PRaVDA collaboration have demonstrated the capability of CMOS imagers to be used as energy-range detectors for pCT, with respect to individual proton imaging [23], and capability of dE/dx measurements at energies relevant for clinical pCT [24]. A further proof of concept has been provided in [25], demonstrating the capability of detecting correlated events in a stack of two CMOS imagers, a condition necessary to perform track reconstruction across a CMOS RT. Additionally, several other authors have demonstrate the capability for CMOS APSs to be used for (pRG) [11] and pCT [12,13].
This paper reports on the development of a large area CMOS APS for pCT and, more generally, for imaging in proton therapy. Optimisation of design parameters, such as epitaxial layer and resistivity, as well as performance in individual proton imaging and proton radiography are reported.

The PRaVDA CMOS Active Pixel Sensor
The PRaVDA CMOS APS has been manufactured in a 0.18µm commercial CMOS process using reticule stitching technique [18,26]. Each sensor offers an imaging area of approximately 5×10 cm 2 with a 198 µm pixel pitch and it is two-side buttable to allow larger imaging areas. For the PRaVDA experiments, two sensors were tiled together to offer an imaging area of approximately 10×10 cm 2 , and a synchronised rolling shutter readout of both sensors was implemented.
Each pixel of the PRaVDA sensor is fitted with 5 diodes: 4 placed at the corner of pixel and one in the middle to improve charge collection efficiency (see figure 1(a)). A conventional readout architecture, shown in figure 1(b), is adopted for this sensor. Pixel values are addressed row-wise via a vertical decoder, enabled by a shift register. For each addressed row, pixel values are sampled along Sample and Hold column stages. Analog pixel values are then converted into the digital domain by an internal single-slope column-parallel Analog to Digital Converters (ADCs) programmable over the range 8-16 bits, loaded onto shift registers and finally clocked out through a number of CMOS digital lines.
Full frame readout of the full imaging area (10×10 cm 2 ) can reach 1000 fps fps (11-bit readout, ADCs clocked at 100 MHz), while higher bit resolution or slower ADC settings lead to lower frame rates, although offering a reduced noise level. A faster readout is achievable when reading out only a fraction of the imaging area (Region of Interest mode).
-2 - During the development phase of the PRaVDA sensors, three combination of epitaxial layer thickness and resistivity were tested in order to reach an optimal trade-off in terms of charge collection efficiency and cluster size, and ultimately image quality.
Specification parameters for the three sensors tested are reported in table 1, including epitaxial thickness, resistivity and charge carrier lifetime (for both epitaxial layer and substrate). Values of resistivity and carrier lifetime are provided by the foundry. The former results from doping concentration specified in the CMOS process used, while the latter is derived from the Technology Computer Aided Design (TCAD) simulations.

Electro-optical performance
Standard electro-optical characterisation for the three sensors reported in table 1 was performed following the EMVA 1288 standard [27]. Mean-variance and linearity curves are show in figure 2(a) and (b), respectively. Noise, conversion gain and Full Well Capacity (FWC) were calculated from the mean-variance curves of figure 2(a) and are reported in table 2. Quantum Efficiency (QE) to green light (523 nm) is calculated from the linearity curves of figure 2(b), after conversion of signal value from Digital Numbers (DN) to e − through the previously measured conversion gain, and is reported in table 2 for the three sensors. Electro-optical parameters reported in table 2 suggest similar performance of the three sensors in terms of noise, gain and FWC. However, the W1 sensor appears to have a significantly lower QE (24%) compared to the other two sensors (51% for W3 and 59% for W5). This difference can be explained by accounting for the lower resistivity of the epitaxial layer reported for the sensor W1 in table 1. Sensor W1 is excluded from further tests, due to its low QE.

The DynAMITe detector
A CMOS Active Pixel Sensor, named DynAMITe, whose performance for proton imaging has been extensively studied both with experimental data [23][24][25] and simulations [29], has been used to compare the response of the PRaVDA CMOS sensors for individual proton imaging. Details of the sensor architecture, electro-optical performance and radiation hardness are reported elsewhere [19,22,28]. Electro-optical performance parameters for this sensor are reported in table 2, for comparison with the PRaVDA CMOS APSs. For the purposes of this work, note that the 50 µm pixel resolution was used and detector was readout in ROI mode, to allow for faster readout. Epitaxial layer for this sensor is 12 µm. Images acquired for this experiment were dark-corrected by subtraction of the average of a number of dark frames, and, subsequently, thresholded with respect to a reference value equal to three times the noise level. A clustering algorithm was used to account for single hit events spread over multiple pixels.

-4 -
Signal spectra for sensors W3 and W5 were measured, as sum of the signal generated in each pixel of a cluster. Cluster size distribution was also evaluated.
To compare the detection performance of sensors W3 and W5 wih the DynAMITe APS, whose response to individual protons is used here as benchmark, sensors were exposed to protons over a range of energies (6-29 MeV). The pristine 29 MeV was degraded to lower energies by insertion of PMMA absorbers at the beam nozzle. Same experimental and image processing procedures as described previously were used. It is to note, however, that a small ROI (10 rows) was used for the DynAMITe detector, given the lower readout speed. Most Probable Signal and average cluster size were assessed from this experiment.

Radiographs
Sensor W3 was used to produce proton radiography images at iThemba LABS (South-Africa) with 60 MeV protons, the average energy expected after the patient in proton CT and radiography.
A patient collimator (see figure 5(a)) was used to assess the capability of the sensor to image complex shapes, in a clinical scenario. The collimator was made of a 6-cm thick brass ring with a 12 cm diameter, and an inner part made of of a 3-cm thick Cerrobend alloy with an internal diameter of 7.5 cm, shaped to match anatomical features. The collimator was placed at the beam nozzle and the sensor at the iso-centre in the proton therapy vault. A full frame readout (95 fps) was used for this experiment.
An imaging phantom was also used to assess the imaging capability of the sensor W3. The imaging phantom features a stepped region, different-size holes as well as different-thickness holes. Phantom specifications are schematically shown in figure 5(b). The phantom was placed in close contact with the detector, and both positioned at the patient position in the proton therapy setting.

Individual proton detection
Normalised signal spectra, in units of DN, measured for the W3 and W5 sensors, exposed to 29 MeV protons, are shown in figure 3(a). For both sensors spectra resemble spread-out Landau energy-loss curves [30], with the addition of low-signal noise which could be either due to external noise sources (e.g. secondary particles) or to "split events", i.e. hits which fall partially outside the small readout ROI (54 rows). Low-energy tails have been excluded from subsequent analysis by applying a threshold at 100 DN. Mean detected signal is 364 DN for W3 and 551 DN for W5, suggesting a different QE and Charge Collection Efficiency (CCE) for the two sensors.
Cluster size distributions are shown for both sensors in figure 3(b). Mean cluster size is 3.3 pixel for sensor W3 and 9.1 pixel for sensor W5.

Comparison with the DynAMITe detector
The response of the PRaVDA sensors W3 and W5 and the DynAMITe detector to protons in the energy range 6-29 MeV is reported in figure 4. Figure 4(a) shows the Most Probable Signal, obtained from Landau fitting of the signal spectra and converted in unit of e − , was plotted as function of proton energy for the three sensors under study. While the response of sensor W3 and DynAMITe is comparable, sensor W5 shows a much increased signal for the lowest energy. Figure 4(b) shows average cluster size as a function of beam energy. While W3 and DynAMITe exhibits a comparable cluster size (in terms of number of pixels), W5 shows larger clusters.
Differences in the response of the three sensors will be discussed in section 5.

Planar imaging
Radiography of a patient collimator is shown in figure 5(a), together with the collimator used and described in section 3. The complex shape of the collimator appears to be correctly reproduced. In figure 5(b), a radiography of the imaging phantom described in section 3 shows the main features of the phantom (stepped regions, holes) as well as the 'halo' effect due to Multiple Coulomb Scattering and discussed in [23]. Although both radiographs show the potential for the PRaVDA sensor W3 to be used for proton radiography, some artefacts such as a horizontal misalignment between the two sensor halves tiled together, a row at the tiling position and a different responsivity of both sensor halves are evident. This makes the need for further image correction techniques for this sensor to be used in proton radiography.

Discussion
Three sensors (W1, W3 and W5) were designed and manufactured by the PRaVDA collaboration with different epitaxial thickness and resistivity (see table 1), with the aim to find the optimal specifications for individual proton imaging. After optical characterisation, one sensor (W1) was excluded from further studies due to a low QE, while the other two sensors appeared to show electro-optical performance comparable with the required specifications.
Although both sensors W3 and W5 showed capabilities of individual proton imaging (see figure 3), some important differences in terms of signal spectra and cluster size were found.
Deposited charge, collected charge and Charge Collection Efficiency (CCE) and average cluster size for 29 MeV protons (W3 and W5) and 27.3 MeV protons (DynAMITe) are reported in table 3, together with other relevant design parameters such as epitaxial thickness and pixel pitch. Deposited charge was calculated using the NIST PSTAR reference database [31], collected charge corresponds -7 - to the Most Probable Signal obtained from Landau fitting of the signal spectra for a given energy (see section 4) and CCE is the ratio of these two quantities. Uncertainty on the Most Probable Signal is smaller than 0.6% for all three sensors. CCE is comparable for DynAMITe and W5 (≈ 90%), while sensor W3 shows a CCE > 1 suggesting than more charge than that deposited in the epitaxial layer is collected, likely due to charge collection from the substrate. Charge collection from the substrate depends on substrate resistivity and thus minority charge carrier lifetime, which is comparable for W3 and Dynamite (≈ 0.048 µs). Charge collection from the substrate also depends on the doping profile across the wafer, i.e. how sharp is the transition from the high doping region of the substrate to the low doping -8 - region of the epitaxial. However, it is worth nothing that a moderate degree of charge collection from the substrate does not represent an issue for proton imaging, while it would be for other applications (e.g. X-ray imaging) due to the different nature of the interaction.
Cluster size provides information on charge sharing. Mean cluster size for DynAMITe and W3 (≈ 3 pixel) is comparable, although DynAMITe has a smaller pitch (50 µm) than W3 (198 µm) effectively resulting in larger clusters. However, given the similar pixel design for the two sensors (i.e. diodes placed at the pixel corners), the similar number of pixels suggests that energy deposited in a single pixel is then shared by nearby diodes of adjacent pixels. For W5, on the other hand, the cluster size is larger (9.1 pixel) and this can be related to a higher degree of charge diffusion in the thicker epitaxial (24 µm) resulting in a larger charge spread before collection, and a longer charge carrier lifetime due to the higher resistivity (1000 Ohm cm) allowing for charge spread at larger distance from the diodes to be collected.
The large cluster size measured for W5 can limit detection performance for individual proton imaging. In fact, such large clusters, expected to become even larger at lower energies, entail pile-up issues. For this reason, only sensor W3 has been included in further testing related to planar integrated imaging.
Planar imaging showed the capability for the PRaVDA APS to be used for proton radiography, although image correction algorithms need to be employed to reduce variation in response between the two sensor halves as well as correcting for artefacts showing at the interfaces between the two halves.

Conclusions
A novel high-speed large-area CMOS APS for imaging in proton therapy has been presented together with initial results. The PRaVDA CMOS sensor is the first prototype offering a large imaging area (5×10 cm 2 ) with a fast readout (kHz) The optimisation of such sensors, in terms of modifying epithickness and resistivity, has been discussed and comparative results provided. Though the radiation hardness of conventional CMOS sensors will always be lower than SSDs, they can through careful design and fabrication still provide a long and economical viable operating life [22]. Though operating speeds are several magnitudes lower than for SSDs, their pixel-based architectures permit much higher number of detected protons per readout cycle. Coupled with the ease of recovering accurate event locations, APS are a practicable sensor for future instrument designs.