Background
Positron emission tomography (PET) is a molecular imaging technique used for the visual and quantitative assessment of a tissue of interest via administration of a radiopharmaceutical tracer. The validation of a new tracer can be achieved by comparing the tracer uptake seen in PET images with the gold standard of histopathology imaging. Once validated, the new tracer can then be used to provide information about diagnosis, prognosis and response to treatment. Due to the heterogenous nature of tumours and the consequent heterogenous distribution of tracer uptake [
1], the validation of PET against histopathology images can only be considered meaningful if spatially corresponding regions are compared.
The process of spatially aligning multi-modal image data sets is challenging and is usually achieved using anatomical landmarks, features and/or pixel intensity values to establish correspondences between images acquired with different modalities. The registration between in vivo human tomographic images and three-dimensional (3D) histopathology data in the brain has been shown to obtain satisfactory results with submillimetre accuracy [
2,
3]. However, the process of obtaining 3D reconstructions of the whole histopathology specimen is laborious, expensive and not feasible in most routine pathology settings. The post processing for such an approach includes (a) registration of individual histopathology slices to block-face images (photographic images of the specimen taken before and after cutting) to correct for the shrinkage incurred during histopathology procedures [
4], (b) registration of shrinkage-corrected histopathology slices to each other to reconstruct a 3D volume [
5,
6] and (c) registration of the histopathology volume to the in vivo imaging data (i.e. in vivo PET/computed tomography (CT)/MR) using intensity-based or other algorithms directly [
2]. In some cases, the in vivo imaging and histology volume are registered using volumetric block-face images as the registration reference [
3] due to the latter’s rich spatial information content. However, the 3D reconstruction of block-face images can be distorted by the so-called ‘banana effect’ whereby a 3D curved object cannot be reconstructed from cross-sections without any additional information [
7]. It is possible to minimise related registration errors by using an additional intermediate tomographic scan, for example, CT or MRI, of the unsliced specimen (i.e. an ex vivo scan) with inserted fiducial markers that can be visualised on both ex vivo and block-face images to drive the reconstruction of the 3D block-face volume [
4]. Shojaii R. et al. [
8] used a catheter filled with cuttlefish ink and flour as a fiducial marker that was visible in multi-modal images. A number of other markers have also been proposed, for example, cotton thread infused with solution of gadopentetate dimeglumine and blue tissue marking dye [
9], plastic sheaths of catheter needles [
10], cotton thread infused with a solution of Magnevist and blue tissue marking dye [
11] and bronze and stainless-steel paints [
12]. However, most of these markers/studies are MRI specific.
In addition, the use of systematic sectioning of the specimen, to obtain millimetre-thick tissue slices, simplifies the cutting of parallel and consecutive slices [
9], as opposed to using a freehand technique. Mega et al. [
13] registered stained whole-brain sections to pre-mortem PET images in Alzheimer’s disease, however, no registration errors were reported. Edwards et al. [
14] presented a system for PET-to-histology registration for image-guided surgery but did not report any registration errors. A few other studies involved the registration of histology with autoradiographs in preclinical settings [
3,
15‐
18] and/or histology with MRI [
2,
9,
12,
19‐
21].
Parallel sectioning of the specimen provides good thickness estimates of the sectioned slices, contiguous sectioning aids in minimising the errors in reconstructing the 3D histopathology volume and whole-tissue histopathology slices provide confidence in matching boundaries with block-face/in vivo/ex vivo data. However, registration of non-parallel, non-contiguously cut and non-whole-sized histopathology slices to, for example, tomographic PET slices is more difficult, particularly in regard to identifying corresponding slices in the axial (i.e. perpendicular to the cutting plane) direction. Because the PET and CT data are inherently 3D and the histology slices are inherently two dimensional (2D), an important aspect of the PET and ex vivo CT to histology registration is the identification of corresponding z-axis levels. The difficulty in identifying the z-axis accurately emerges from the lack of a 3D reconstruction of the specimen volume after slicing. Although the specimen is cut into consecutive slices, standard pathology procedures do not involve the use of a system that ensures all slices are cut to exactly the same thickness. As a consequence, a histology slice might not be cut perpendicular to the z-axis of the specimen and different areas of the micron-thick histology data could belong to different millimetre-thick slices.
Therefore, the aim of this work was to develop a registration methodology for an accurate spatial alignment between PET/CT images of the head and neck region in humans and histopathology slices obtained in routine pathology settings where slices may be not perfectly parallel, non-contiguously cut and non-whole-slice sized.
Discussion
We have developed and assessed a method for alignment of PET and histopathology slices obtained in routine pathology settings in six male larynxes with advanced HNSCC cancer. The average overall registration error between PET and histopathology was 3.0 ± 0.7 mm2 which is better than the 6.00-mm FWHM spatial resolution of the PET scanner.
One of the most challenging aspects of the process and one that future efforts should be focused on is the identification of the corresponding
z-axis levels between the 2D histology and the 3D imaging data. When parallel consecutive slices are used, the slice thicknesses are accurately known and the identification of the
z-axis can be performed by simple counting [
27]. However, not all pathology departments have access to the equipment required, and the processing of these specimens would have to deviate from the local pathology laboratory routine. Contrary to previous attempts where the pathology routine had to be changed to introduce either whole-mount tissue sections or parallel consecutive slicing [
27], the method outlined in this study affects the routine pathology workflow with the addition of one extra step, the insertion of the fiducial markers. This step, however, does not require extra personnel as it is not technically challenging and it can be performed easily by a member of the pathology team handling the specimen. In addition, it extends a procedure that has a total duration of many hours (from specimen removal after surgery to generation of tumour blocks for histopathology analysis) by few extra minutes only. For most routine pathology departments in which these kinds of study are performed, the pathologist performs parallel cutting with the use of a bandsaw, and the accuracy of this process depends on the experience of the pathologist and the tissue characteristics. With the use of optical images and the recorded anatomical information from the pathology records, we were able to identify the approximate level of each one of the tissue blocks on the ex vivo CT.
We employed sea urchin spines as fiducial markers which were clearly visible on both CT and histology images. The use of spines can add value for cases where bone tissue is absent.
‘Block-face’ images were obtained on both sides of the tissue blocks since tissues used for H&E staining were extracted from either side of the blocks and these images acted as a visual reference during registration. Even though block-face images are mainly used for shrinkage correction [
4], the choice of rigid registration for this study obviated the exclusion of a histology-to-block-face step. However, as opposed to other organs, the lung for example [
28], only small deformations occur inside the laryngeal cartilage area. An average shrinkage of 3 % before and after fixation has been calculated for the area included by the cartilage skeleton, which in most cases includes the tumour bulk [
4]. This is because the larynx is an anatomical site with a strong and rigid skeleton, which helps to maintain the shape of the tissues inside it. In addition, it has been shown that the main shrinkage in the whole procedure occurs during the processing of the histological material to acquire the H&E sections from the thick slices and not from the fixation process [
4]. Finally, excluding the use of block-face images avoided potential additional sources of error arising from (1) 3D volume reconstruction of block-face slices, (2) ex vivo CT to block-face registration and (3) histology to block-face registration.
Identification of reliable landmarks is challenging on PET, and therefore, the use of a PET-CT scanner was a major advantage. This is because an intermediate high-resolution ex vivo CT of the larynx specimen served as the reference data set that corresponded well with the in vivo CT data and also with the histopathology slice boundaries and/or edges. A disadvantage was the error due to PET-CT misalignment which was reported from literature and not measured, though the error introduced is likely to be small. A misalignment of ~1 mm may be expected in a worst case spatial calibration between PET and CT of the PET-CT scanner [
29]. It was not possible to measure the PET-CT misalignment explicitly in this study; however, each PET-CT data set was visually verified by an expert clinician, and no further adjustment was deemed necessary in most of the data sets. Somer et al. [
29] report an average PET-CT misalignment due to patient motion during the PET-CT scan as 5.4 mm; however, this figure was obtained primarily from measurement on the body and limbs and so is likely to greatly overestimate misalignments in the head and neck region. Nevertheless, further investigation of PET-CT misalignment would be useful, and the use of a rigid fixation device (e.g. a radiotherapy mask) may be advantageous.
Caldas-Magalhaes et al. [
4] reported an average (in-plane) registration RMSE of 3.49 mm between PET and histopathology in laryngeal and hypopharyngeal cancer using a semi-automatic methodology, i.e. very similar to our overall error of 3.0 mm. Their histopathology samples were whole-mounted histology slices, obtained from a contiguously cut tissue block immersed into agarose medium whilst a meat-slicing machine was used for cutting. None of these methods are part of the routine pathology practice where usually only 30 × 21 mm slides are used and the specimen is sliced freehand without the provision of a supportive structure, such as the agarose medium. The carbon rods used as fiducial markers were not visible on histology, and therefore, manual landmarks were identified on the cartilage to drive the histology-to-block-face image registration. A study of the prostate by Park et al. [
30] reported an average registration error of 7.7 mm between PET and histopathology using an in vivo MRI scan as the reference space. However, obtaining PET, CT and MR in vivo images may not be feasible in every clinical trial. Moreover, this study obtained contiguously cut 3-mm-thick parallel tissue slices with whole-mounted histology slices which again is not a routine pathology practice.
The current study suffers from a number of limitations. In particular, this is a specific case where tumour is surrounded by the larynx cartilage which prevents (in most cases) any unexpected deformation of the tumour during surgery, fixation and the slicing procedure.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
This was a complex study requiring co-ordination, flexibility and perseverance between a large number of contributors. MOD, PKM, AC, MM, TGU and EO were responsible for the study conception and design. RO, RS, JPJ and MM performed the laryngectomies. RHS and EO were responsible for processing of the specimens and histopathology procedures. PB, TP and AC performed the digital microscopic imaging. AC and TGU were responsible for patient recruitment. AC and MOD organised and performed the PET studies. TP, AC, AR, EO and PKM developed the registration methodology. TP and AC performed the registration procedures. TP and AC performed the data analysis. TP, AC and PKM wrote the manuscript. All authors read and approved the final manuscript.