Purpose: We apply positron emission tomography (PET) to elucidate changes in nanocarrier extravasation during the transition from premalignant to malignant cancer, providing insight into the use of imaging to characterize early cancerous lesions and the utility of nanoparticles in early disease.
Experimental Design: Albumin and liposomes were labeled with 64Cu (half-life 12.7 hours), and longitudinal PET and CT imaging studies were conducted in a mouse model of ductal carcinoma in situ. A pharmacokinetic model was applied to estimate the tumor vascular volume and permeability.
Results: From early time points characterized by disseminated hyperproliferation, the enhanced vascular permeability facilitated lesion detection. During disease progression, the vascular volume fraction increased 1.6-fold and the apparent vascular permeability to albumin and liposomes increased ∼2.5-fold to 6.6 × 10−8 and 1.3 × 10−8 cm/s, respectively, with the accumulation of albumin increasing earlier in the disease process. In the malignant tumor, both tracers reached similar mean intratumoral concentrations of ∼6% ID/cc but the distribution of liposomes was more heterogeneous, ranging from 1% to 18% ID/cc compared with 1% to 9% ID/cc for albumin. The tumor-to-muscle ratio was 17.9 ± 8.1 and 7.1 ± 0.5 for liposomes and albumin, respectively, indicating a more specific delivery of liposomes than with albumin.
Conclusions: PET imaging of radiolabeled particles, validated by confocal imaging and histology, detected the transition from premalignant to malignant lesions and effectively quantified the associated changes in vascular permeability. Clin Cancer Res; 17(3); 550–9. ©2010 AACR.
Detection of cancerous progression is required for differential diagnosis and therapeutic planning. Here, the sensitivity and accuracy of pharmacokinetic modeling using positron emission tomography (PET) are exploited to detect the transition from premalignant to malignant breast cancer lesions through a characterization of the vasculature. The tumor vascular volume fraction is first estimated from carrier-conjugated radioactivity immediately after injection; transport into the interstitium is then estimated. Albumin, a 4 × 4 × 14 nm tracer, has been previously applied to characterize normal vascular function as a small increase in the number of gaps through or between cells or the formation of vesicular vacuolar organelles significantly increases albumin permeability. We demonstrate that the apparent permeability of 64Cu-labeled albumin provides a sensitive indication of tumor transition. 64Cu-labeled liposomes augment this measurement by precisely defining the time of transition and by mapping the heterogeneity of permeability.
Sensitive and quantitative measures to characterize cancerous progression are required for differential diagnosis and therapeutic monitoring. Imaging techniques allow us to diagnose and monitor cancer progression longitudinally; positron emission tomography (PET) is a functional imaging technology that provides rapid, reproducible, noninvasive in vivo assessment and quantification of biological processes involved in tumor development. In this study, we detect cancerous progression within a mouse model of ductal carcinoma in situ (DCIS) by using PET to track nanometer-sized tracers. The model recapitulates the clinico-epidemiologic observations in human disease and is referred to as mouse mammary intraepithelial neoplasia outgrowths (MIN-Os), which consistently undergo malignant transformation to a phenotype capable of autonomous ectopic growth (1–3). The MIN-O model allows us to study the pharmacokinetics and tumoral uptake of PET-labeled albumin and liposomes at various points in the developmental timeline of mammary tumors (4). Our aim was to determine the utility of PET to detect the tumor transition through a detection of changes in vascular volume fraction and permeability. While 2-[18F]-fluoro-deoxy-D-glucose (18F-FDG) has previously been applied to follow changes in tumor metabolism (5), PET-based imaging of nano-structures such as liposomes has not previously been applied to follow the malignant transformation, and 18F-FDG was unable to detect the vascular changes associated with the malignant transition (5).
To accurately assess the timing of the transition, we combine measurements obtained with radiolabeled albumin and liposomes. Albumin (∼66 kDa, dimensions of 4 × 4 × 14 nm) has been extensively used for permeability studies (6, 7), and is a versatile protein carrier for targeting of hydrophobic molecules in clinical settings. The preferential localization of albumin in tumors and inflamed tissue due to increased vascular permeability, availability, biodegradability as well as lack of toxicity and immunogenicity also make it a good candidate for drug delivery (8). Liposomes are vesicles composed of 1 or more concentric phospholipid bilayers and are considered to be a prototype of nanocarriers under development (9). The typical diameter of liposomes is 65–120 nm, with the small diameter facilitating extravasation through leaky blood vessels and fusion with cellular membranes (10, 11). Prolonged blood circulation of the liposomes is achieved with the addition of a polyethylene glycol (PEG) coating to resist opsonization, which efficiently minimizes their removal by macrophages (12, 13). Liposomes facilitate controlled release of drugs, thereby reducing drug-related toxicity and facilitating targeted delivery of drugs.
On the way to their target, tracers or drug carriers meet physiological barriers, such as interstitial hypertension and abnormal tumor vasculature (14–16). The altered tumor endothelium includes modified intercellular junctions as well as specialized transendothelial pathways such as vesicular vacuolar organelles (VVOs), fenestrations and transendothelial gaps, often characterized as an effective “pore size cutoff” (17–21). Elevated levels of permeability enhancing factors, including bradykinin, nitric oxide (NO) and vascular endothelial growth factor (VEGF) create these changes and can be therapeutic targets for neutralization (22–25). As a result of the altered endothelial morphology, nanoparticles of the optimal size exhibit enhanced permeability and retention (EPR) within the tumor interstitium (24). Permeability is particularly enhanced for small carriers (i.e., <10 nm) but interstitial retention is more pronounced for larger particles (16). Reduced lymphatic drainage further increases the retention of molecules in the tumor interstitium. Diffusion of larger particles is less than of smaller particles, slowing the redistribution of large particles to the blood circulation. Thus, larger tracers are required to detect and map vascular changes associated with the tumor transition.
To quantify such changes, we evaluate a PET-based pharmacokinetic model (26) and augment PET with confocal imaging. Many noninvasive techniques have been developed to evaluate particle pharmacokinetics (27–40). While the spatial resolution of MRI (magnetic resonance imaging) and CT (computed tomography) is superior to PET, PET has advantages in sensitivity and quantitation, since picomolar concentrations can be detected and estimates of blood pool and organ tracer concentration are accurately mapped in time and space. The 12.7-hour half-life of 64Cu facilitates intratumoral particle tracking over 48 hours; repeated dosing facilitates imaging of longitudinal disease progression over months of disease progression. Thus, we apply 64Cu-labeled albumin and liposomes to track intratumoral delivery over the progression from the premalignant to malignant phenotype and evaluate the ability of PET to characterize disease progression.
Materials and Methods
Details of each procedure are described in the Supplementary methods.
MIN-O is a mouse model of DCIS that recapitulates “premalignant” disease of the mammary gland. The course of disease approximates the neoplastic progression of human breast carcinoma from preinvasive DCIS to invasive carcinoma. Outgrowths transplanted into gland-cleared mammary fat pads of syngeneic immunocompetent mice result in hyperplastic and dysplastic growth, which remains confined to the dimensions of the stroma and is analogous to human DCIS (3).
All animal studies were conducted under a protocol approved by the University of California, Davis Animal Use and Care Committee.
Three studies were conducted (n = 27 total), involving MIN-O subline 4 (tumor latency of 5 weeks) in the first 2 studies and subline D (latency of 7–8 weeks) in the third study. The first group of animals (line 4, n = 8) was PET and CT scanned at 3, 5, 7, and 8 weeks after transplantation to follow the development of tumors longitudinally. A second cohort was implanted with the same tumor line (n = 8) and sacrificed after PET scanning at 5 weeks in order to catch the transformation from premalignant to malignant tissue. A third cohort of animals (line D, n = 11) was imaged with confocal microscopy and PET at 5 and 7 weeks posttransplantation.
Tumor permeability and distribution of PEGylated liposomes and albumin were investigated. Liposomes were prepared as described previously (26, 40, 41) and had an average diameter of 100–115 nm. Fatty acid and globumin-free bovine serum albumin (BSA, A0281) was purchased from Sigma-Aldrich (99%, MW of 67 kDa).
In vivo and ex vivo imaging was performed with a microPET scanner (Focus 120, Siemens Medical Solutions, Inc). Animals that received radiolabeled liposomes were imaged at 0, 6, 18, 28, and 48 hours after tail-vein injection and data were acquired for 30 minutes. Due to the shorter half-life of albumin, animals receiving 64Cu-BSA were scanned at 0–2, 6, and 18 hours after injection (up to 48 hours for cohort 2), where data were acquired for 2 hours and rebinned into a dynamic scan for the 0–2 hour time point.
To investigate the colocalization of vessels and tracers, animals were injected through the tail vein with Alexa-555 conjugated tracers and lectin for vessel identification (Fluorescein Lycopersicon esculentum (tomato) lectin from Vector laboratories [2 mg/mL, 50 μl]).
The tumor tracer concentration, transport constant, λ (h−1), vascular volume fraction (%), and apparent permeability (AP, cm/s) were estimated based on the time activity curves (TACs) for each week and for each tumor using an image-driven pharmacological model developed by Qin et al., 2009 (26).
After implantation of the hyperplastic lesions, thin layers of hyper-proliferative cells expanded through the mammary fat pad, inducing substantial angiogenesis prior to tumor formation (Supplementary Fig. S1). At week 3, even small nonpalpable lesions were detected by PET with both albumin and liposomal tracers (8/16 lesions were palpable but all were detected by PET). At 5 weeks after implantation, solid tumors were observed at scattered foci, with a fibrous stroma replacing the fat pad in the tumor vicinity (Fig. 1A and B). Regions with high intensity on ex vivo PET images corresponded geometrically with regions of dense outgrowth or tumor tissue (Fig. 1A and B) and were also readily visualized with in vivo imaging (Fig. 1C). Even at this early time point after transplantation, the accumulation for both albumin and liposomes was greater in the lesion than in the healthy fat pads and surrounding muscle, where the accumulation of the tracers did not exceed the background level (Fig. 1C). At 5 weeks after implantation and 48 hours after injection, the average tumor accumulation across the 2 tracers was ∼4.3%ID/g (with albumin accumulation exceeding liposomal accumulation) and was significantly greater than the accumulation of 0.6%ID/g in surrounding striated muscle tissue (P < 0.05, ANOVA). Muscle and tumor accumulation was similar between PET and biodistribution data (Fig. 1D), and ex vivo and in vivo PET estimates of radioactivity in identical breast tumors and fat pads were correlated (R2 = 0.95, least square analysis).
In vivo PET imaging was then applied to visualize and quantify vascular density and permeability longitudinally over tumor development (Fig. 2). Serial imaging of a group of animals was conducted as the tumors developed from the premalignant to malignant state (Fig. 2A) over weeks 3, 5, 7, and 8 after implantation, alternating the injection of liposomal (Fig. 2A panels ii and iv) and albumin (panels i and iii) tracers. In the same animal, as the tumors progressed, the accumulation of both tracers increased as detected by increased PET signal intensity in the lesions (Fig. 2A).
Image-based estimates of the tumor volume increased rapidly starting around 5 weeks after transplantation (Fig. 2B), indicating a transition from a premalignant to a malignant phenotype at this time point. The tumor doubling time decreased as the tumors progressed; the mean tumor doubling time pretransition (3–5 weeks), at the transition (5–7 weeks), and posttransition (7–8 weeks) was 13.3 ± 6.8, 11.4 ± 6.4 and 8.4 ± 4.5 days, respectively (P = 0.13, ANOVA between pre- and posttransition, Fig. 2C). The mean tumor doubling time for the entire time period was 9.5 ± 2.8 days. PET images of the liposomal (Fig. 2D upper) and albumin (Fig. 2D lower) tracer distribution correlated with histology, where regions of necrosis were visible. Cystic regions were prominent in the majority of the tumors at the 8-week time point. The tumor mass detected on CT images corresponded with regions of increased intensity on PET images for both tracers (Fig. 2D). By the 8-week time point, regions of dense tumor tissue were observed with extensive vasculature (Fig. 2D).
Albumin uptake precedes liposomal uptake in tumor transition
Lesion growth and progression were reflected in the tumoral tracer uptake, which increased as the transition occurred (Fig. 3). Liposomes had a longer half-life in blood than albumin (i.e., ∼18 and ∼6 hours, respectively) thus enabling accumulation of the larger liposomal tracer over a longer period of time (Fig. 3A); in each case scanning continued until the blood concentration was less than 10% ID/cc. After removing the estimated blood pool concentration, the tumor accumulation of liposomes demonstrated a stepwise transition between the premalignant (weeks 3–5) and malignant (weeks 7–8) state, with the temporal peak accumulation increasing up to 3-fold between weeks 3–5 and 7–8 (Fig. 3B).
The transition in the accumulation of the much smaller albumin molecule was more gradual, where differences in peak accumulation were significant between weeks 3–5, 3–7, 3–8, and 7–8 (Fig. 3C, P < 0.05, ANOVA). Comparing the TACs of the 2 tracers, the time to peak accumulation was shorter for albumin; in week 8, accumulation within the first hour was as high as 3% ID/cc (Fig. 3B and C) and accumulation of 3% ID/cc required 6 hours for liposomes. The rate of clearance of the tracers over 48 hours was similar. At week 5 after implantation, 25.5 ±16.9% of the liposomal tracer had cleared from tumors by 48 hours, as compared with 33.6 ± 9.9% of the albumin tracer (data not shown, P = 0.06, ANOVA).
The tumor concentration of the tracer at the time of the latest scan (i.e., 48 hours for liposome injection and 18 hours for albumin injection) reached ∼6% ID/cc at week 8 and was confirmed by the biodistribution (Fig. 3B–D). The tumor-to-muscle ratio was significantly higher with the liposomal tracer, that is, 17.9 ± 8.1 and 7.1 ± 0.5 for liposomes and albumin respectively, indicating a more specific delivery of liposomes than with albumin. In general, liposomal radioactivity and the radioactive metabolites were concentrated in the spleen and intestines, while radioactivity was greater in liver, kidney, and muscle tissues after the injection of 64Cu-labeled albumin. The accumulation in blood, heart, and lungs was similar between the 2 tracers.
PET facilitates direct visualization of tumor heterogeneity
Imaging the same animal at weeks 3 and 7 or 5 and 8, mean accumulation in the same tumors increased 1.6- and 2.3-fold, respectively, for the albumin and liposomal tracers (P <0.02 paired t-test, Fig. 4A). The initial accumulation in week 3 and final accumulation in week 8 was similar for the 2 tracers. Thus, for liposomes, accumulation increased more rapidly at the time of the transition (between weeks 5 and 7) whereas albumin accumulation gradually increased between weeks 3, 5, 7, and 8 (Fig. 4B).
To assess tumor heterogeneity, estimates of maximum and minimum accumulation were generated with a resolution of (0.8 mm)3 and compared with the mean over the tumor (Fig. 4B). Comparing liposomal and albumin accumulation, the spatial maximum and the heterogeneity of liposomal accumulation were larger (Fig. 4), where the spatial variation of liposome accumulation ranged from <1% to ∼18% ID/cc and the spatial variation of albumin accumulation was smaller, ranging from ∼1% to 9% ID/cc. At 8 weeks after implantation, the accumulation of albumin and liposomes (peak over time and average over animals) was 5.7 ± 0.5% and 6.0 ± 1.1% ID/cc, respectively, reflecting the similar mean and greater heterogeneity of liposomal accumulation. While the minimum tumor accumulation observed following liposome injection was smaller than accumulation following albumin injection, the difference was not significant.
PET images also demonstrated the differences in the rate of accumulation and intratumoral heterogeneity of the 2 tracers (Fig. 4C and D). Images and surface plots of a single slice through the tumor center at the18-hour time point 8 weeks after implantation convey spatial differences in tumoral tracer distribution between the 2 tracers. Tumors displayed a heterogeneous pattern of tracer distribution (Fig. 4C), with a minimum in the tumor center in which activity accumulated slowly. The high intratumoral heterogeneity of the liposomal tracer was reflected by a larger spatial variance in intensity, with high accumulation peripherally and large central subregions with lower accumulation. The rapid accumulation of the albumin tracer was also clearly observed (Fig. 4D i and iii), indicating a more rapid extravasation of the smaller-sized tracer into the tumor interstitium. Across the tumor, the spatial variation in albumin accumulation was less than that observed for liposomes (Fig. 4D iv). For example, the mean accumulation of albumin and liposomes in these central slices was 5.1 ± 1.5% and 5.4 ± 3.9% ID/cc, respectively.
Vascular volume fraction and permeability increase with tumor transition
The tumor vascular volume fraction (estimated from tumor radioactivity immediately after injection) increased from 5% to nearly 8% during the progression from the premalignant to malignant phenotype (Fig. 5A) and was similar for the liposomal and albumin tracers. CD31 images confirmed an extensive vasculature and the development of large supplying arteries in weeks 7 and 8 (Fig. 2A). The vascular volume fraction was also estimated based on the CD31-positive endothelium and increased from ∼3% to ∼8% over the transition (Fig. 5B), with an increase in median vessel diameter from 10 to 15 μm, respectively. Segmenting out subregions within the tumor revealed intratumoral heterogeneous perfusion, with vascular volume subfractions reaching 10% and 16% in highly vascularized areas in albumin and liposomes tumors, respectively (data not shown).
The apparent vascular permeability increased approximately 2-fold during the transition from the premalignant to malignant phenotype, peaking near 1.3 × 10−8 cm/s for liposomes and 6.6 × 10−8 cm/s for albumin (Fig. 5C). The change in AP to albumin was gradual, with an increase of ∼1 × 10−8 cm/s between weeks 3–5 and 7–8 and a larger increase of 2.8 × 10−8 cm/s between weeks 5 and 7 (Fig. 5D). Permeability to liposomes increased in a stepwise fashion, with a substantial change observed only between weeks 5 and 7 (Fig. 5D). There was also large intratumor heterogeneity in permeability, where the variation in AP ranged between 2- to 3.8-fold for albumin and 1.2- to 1.6-fold for liposomes.
Confocal imaging of tracers validates kinetics and accumulation at the time of transition
To visualize the microscopic distribution of the tracers in ex vivo tumor interstitium, fluorescent albumin and liposomes were imaged together with a vascular lectin after circulation for 0–28 hours. Immediately after injection of the labeled liposomes (i.e., 30 minutes), minimal, isolated fluorescence was observed; however, at 18 and 28 hours, particle fluorescence had increased with colocalization of vessel walls (in green) and liposomes (Fig. 6A). With fluorescent albumin, immediately after injection, the tracer was concentrated near the tumor vasculature, whereas at 18 hours, the tracer distribution was homogenous within the tumor interstitium (Fig. 6B). By 28 hours, the majority of the remaining tracer was detected near the tumor vessels. The dynamics of the tracers are visualized in Fig. 6 C and D, where the measurements within the ROIs placed adjacent to vessels differed for the 2 tracers, indicating that the albumin distribution within the tumors is dynamic, clearing more rapidly than the liposomes. With both tracers, particle fluorescence was much lower in healthy muscle than in the tumors (Fig. 6A and B).
Quantitative measurement of biochemical processes underlies progress in cancer biology and treatment. Here, we demonstrate that in vivo PET imaging can detect the transformation from a premalignant to malignant breast cancer by quantifying the extravasated tumor concentration of tracers and the vascular volume fraction and blood level of tracer at each time point. Small (∼mm3) lesions were detected by microPET using 64Cu-based tracers and correlated with foci of early disease. The spatial resolution of the microPET scanner used in this study was <2 mm and was sufficient to demonstrate to detect and map accumulation in this model; similar resolution is possible in future human studies as well. While other modalities have advantages in spatial and temporal resolution, the sensitivity of PET to picomolar concentrations of material and its inherent quantitation of pharmacokinetic parameters suggest that PET is an obvious choice for such studies. PET estimates of changes in both vascular volume fraction and permeability were straightforward and robust; such estimates of vascular and extravascular kinetics are problematic with other imaging methods.
Leveraging the understanding of albumin exchange in normal and other hyperpermeable vascular beds, we employ an albumin tracer to detect early changes in permeability in the MIN-O model of DCIS. The MIN-O lines provide insight into the origin, evolution, and outcome of disease (1, 2, 5, 42). Accumulation of the tracers was minimal in normal fat pads and surrounding striated muscle, validating that the enhanced uptake in the transplanted MIN-Os was due to progressing disease. We found that the vascular volume fraction increased from 3% to 8% and the apparent vascular permeability to albumin and liposomes increased ∼2-fold to 6.6 and 1.3 × 10−8 cm/s, respectively, over the transition.
Accumulation of the 4 × 4 × 14 nm albumin tracer increased earlier in the disease progression than the 100-nm liposome tracer and was more gradual, although ultimately the accumulation was similar in the fully malignant lesions (Fig. 4B). An increase in accumulation can be accounted for in terms of increased convection and diffusion through modified intercellular junctions as well as specialized transendothelial pathways such as VVOs and transendothelial gaps. These mechanisms can contribute to both albumin and liposome transport but their relative contributions in different tumors are controversial (20). Although our measurements of AP do not discriminate between these mechanisms, we emphasize the effectiveness of the albumin tracer to detect early changes in permeability. Specifically, albumin normally crosses the vascular permeability barrier via restricted penetration through “small pores” whose size is determined by junctional proteins and matrix structures within the intercellular junctions (21). Albumin, with an AP of 1 × 10−8 cm/s in normal muscle, represents the upper limit of normal permeability and increases as a result of an increased number of larger pores such as those associated with gaps through or between adjacent cells or the formation of VVOs. These changes can be detected with a tracer size of albumin because a small increase in the number of such large pores significantly increases albumin permeability. An increase in the accumulation of lower molecular weight tracers can result from an increase in the number of vessels, an increase in the number of small pores or a transition to larger pores; smaller tracers are therefore both less specific and less sensitive to changes in the number of large pores. Although the small molecule tracer 18F-FDG detected the progressive increase in tumor metabolism in a previous study (5), changes in vascular volume fraction and permeability were detected only with the larger probes employed here. Furthermore, albumin is still small enough to easily penetrate pores that may restrict tracers as large as liposomes.
The early increase in albumin accumulation (Fig. 3C, week 3–5 and Fig. 5D) before liposomal accumulation increased (Fig. 3B, week 5–7 and Fig. 5D) and the subsequent increase in permeability to both tracers is consistent with the formation and increase in frequency of a large pore pathway; however, the contribution of specific mechanisms involving VVOs, transcellular gaps, or the redistribution of proteins controlling intercellular junction permeability cannot be determined (20). Similarly, the initial transition is indicated by the increase in albumin permeability without as large a change in liposomal permeability.
The above argument emphasizes the effectiveness of albumin (or a tracer with similar size and properties) as a tracer to follow tumor progression. Additional factors influence the net delivery of tracers of different size. The blood half-life of the albumin and liposomal tracers applied here were 6 and 18 hours, respectively, and the improved permeability of the albumin was balanced by the loss in circulation time. The tumor-to-muscle ratio was significantly improved with liposomes as compared with albumin, limiting off-target effects.
Further, a comparison of the accumulation of the 2 tracers emphasized the heterogeneity in the effective permeability size cutoff. By week 8, both tracers reached similar mean intratumoral concentrations of ∼6% ID/cc (Fig. 4B) but the distribution of liposomes was more heterogeneous reaching concentrations as high as 18% ID/cc (as compared to a 9% ID/cc peak concentration of albumin; Fig. 4B). High IFP and low diffusion and convection will influence the penetration depths of liposomes, as reported in multicellular tumor spheroids (43, 44) and in experimental solid tumors (45) and thus the relative distribution of the tracers is likely to indicate the distribution of large pores. PET is an effective method to map the relative concentration of tracers over time and space, where maximum–intensity projection images and contour maps were each useful in visualizing variations in accumulation. Due to this tumoral heterogeneity, fully treating a large tumor with a 100-nm particle may be difficult as a larger dose is required to reach a minimum concentration throughout the tumor. Thus, treatment of in situ disease could be facilitated by the development of long-circulating carriers with a diameter on the order of tens of nanometers. There is a substantial literature detailing differences in the accumulation of radiolabeled antibodies (46); such studies will be important for new nanotechnologies as well.
In summary, we find that a 4 × 4 × 14 nm radiotracer provided a sensitive probe to detect the progression from a premalignant to malignant lesion and facilitated a quantitative estimate of vascular volume fraction and apparent vascular permeability. In addition, mapping enhanced permeability to a 100-nm liposomal tracer characterized the timing of the transition while also mapping the spatial heterogeneity of the size-dependent permeability. PET, as employed here, can quantitatively assess tumor progression and inform choices in drug vehicles.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
We acknowledge the following assistance: Katie Bell (immunohistochemistry), Dave Kukis (radiolabeling of albumin), Jennifer Fung and Douglas Rowland (imaging), Jinxiu Cheng-Liao (image reconstruction), and Mario Hlawitschka (image segmentation).
NIHR01CA103828, NIHR01CA134659 and the Western Norway Regional Health Authority (Helse Vest).