Marianna Inglese1,2, Matteo Ferrante1, Tommaso Boccato1, Shah Islam3, Matthew Williams4,5, Adam D Waldman6, Kevin O'Neill7, Eric O Aboagye3, and Nicola Toschi1,8
1Biomedicine and Prevention, University of Rome Tor Vergata, Rome, Italy, 2Surgery and Cancer, Imperial College London, London, United Kingdom, 3Department of Surgery and Cancer, Imperial College London, London, United Kingdom, 4Computational Oncology Group - Department of Surgery and Cancer, Imperial College London, London, United Kingdom, 5Institute for Global Health Innovation, Imperial College London, London, United Kingdom, 6Centre for Clinical Brain Sciences, University of Edinburgh, Edimburgh, United Kingdom, 7Imperial College Healthcare NHS Trust, London, United Kingdom, 8Department of Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Boston, NY, United States
Synopsis
Keywords: Data Processing, Modelling, Deep Learning, convolutional filters
Stratifying
human brain gliomas using imaging techniques is extremely challenging. Valuable
insight into the characterization and classification of gliomas can be provided
by integrating two imaging modalities, i.e.
18F-FPIA PET and MRI. This
study introduces a new approach for glioma stratification based on the
extraction of temporal features from tissue time activity curves (TACs)
extracted from dynamic PET/MRI data. We exploit tissue-specific biochemical
properties embedded in the TACs through deep learning and achieve good
discrimination results while foregoing pharmacokinetic fitting and hence invasive
measurement of the AIF.
Introduction
Gliomas are the most common primary brain tumours, and there is wide interest in imaging
techniques able to stratify such lesions1. 18F-fluoropivalate
(FPIA) PET/MRI integrates two imaging modalities that can help the
discrimination of brain gliomas2. Both PET and MRI
dynamic acquisitions provide maps of the spatiotemporal concentration of the
tracer/contrast agent in vivo, conveying information about the delivery
of the compound to tissue, its interaction with the target as well as its
washout. Therefore, tissue time activity curves (TACs), along with the
information about the concentration of the agent in the plasma, are typically employed
for pharmacokinetic fitting. The primary disadvantage of this strategy is the necessity
for invasive arterial blood sample collection as well as a priori choice
of the pharmacokinetic model, which in turn is tracer dependent and
non-univocous.
We propose an alternative approach to glioma stratification based on the
analysis of dynamic intensity patterns. In particular, we evaluate the
potential of temporal features extracted from multimodal (i.e. dynamic PET and MRI images) TACs in discriminating high-grade gliomas (HGGs) from low-grade gliomas (LGGs) through deep learning models.Materials and Methods
Dataset: 10
patients (4 low (WHO grade II) 6 high-grade (WHO grade III, IV) gliomas; age 31
to 79, mean (+/-SD) 59 ±15.9) were recruited for this study. 18F-FPIA
PET/MR images were acquired on a Signa PET/MR scanner (GE Healthcare). For each
patient, an average of 25202 (±14337)
TACs were extracted voxelwise from dynamic contrast-enhanced (DCE), dynamic susceptibility contrast
(DSC) and dynamic 18F-FPIA PET images using a
lesion mask generated manually by an expert radiologist. TACs were linearly resampled
onto a uniform time axis (96 time points) and standardized.
Model and Implementation: The model consisted in three mono-dimensional
convolutional layers using a gelu, tanh and relu activation
functions, respectively. The filter size was set to 8, 16 and 32, the kernel
sizes to 4 and the strides to 5. The output of the last convolutional layer was
flattened into one fully connected layer with 64 neurons using a relu activation function followed by 0.25 Dropout layer and the last softmax
layer for classification (Figure 1). All experiments were conducted using
Python version 3.8, the Keras deep learning library, using TensorFlow as the
backend. We employed a Linux machine and two Nvidia Pascal TITAN X graphics
cards with 12 GB RAM each. The data was split into training and test sets
(80/20), and hyperparameter optimization within the train set was performed in
a threefold cross validation fashion. An
early stopping method was used to select the optimum number of training epochs
and the batch size (Keras callback function monitoring the loss function with
patience set to 10). Performances were evaluated on test set using both mono- and multimodal TACs to assess
the contribution of each imaging modality. Results
Multimodal
and monomodal TACs (from one single modality at a time, e.g. PET, DCE and DSC)
were used to train the model multiple times, validation results are summarized
in Table 1. Performances are expressed in terms of area under the receiver
operating characteristic curve (AUC), accuracy, precision and recall. The AUCs
are also plotted in Figure 1. Comparable performances were obtained from both
single and multimodal TAC
classification, except for DCE-TACs which delivered the highest accuracy (88%)
and an AUC of 0.8 (Table 1, Figure 1). The worst performance was
obtained by DSC-TACs which delivered 61% accuracy and 0.53 AUC. Interestingly,
DCE-derived
perfusion parameters obtained fitting five pharmacokinetic models (Toft’s3,
extended Toft’s3, shutter speed4, extended shutter speed5 and non-exchange model)6 and averaged over the whole lesion
were not statistically different between LGG and HGG (p > 0.05, Figure
2).Discussion and Conclusions
Gliomas
are the most frequent primary brain tumours, and stratifying their behaviour is
a growing field of research. 18F-FPIA PET/MRI allows the assessment
of several tumour characteristics such as cell proliferation (using
fluoropivalate)2 whose crucial source of nutrition
is thought to be provided by fatty acids, and their blood brain barrier
integrity (through the evaluation of gadolinium-based contrast agent perfusion
and permeability, i.e. DCE and DSC-MRI sequences). Dynamic PET/MRI acquisitions
contain information about time evolution of the signal and how it is affected
by the biochemical properties of the tissue under investigation. However, full
quantification of dynamic PET/MRI data requires both information about
tracer/contrast agent concentration in plasma (AIF), and a priori election of
the kinetic model. In turn, this can affect the reproducibility and reliability
of pharmacokinetic studies, which often show discordant results.
Our pilot
study, even if limited by the small sample size, introduces a promising new approach
for the classification of tumor grade based on time series processing7, hence overcoming the challenge of
pharmacokinetic model fitting6,8, which in the case of PET often includes
arterial cannulation for AIF calculation8.Acknowledgements
All authors declare that they have no known conflicts of
interest in terms of competing financial interests or personal relationships
that could have an influence or are relevant to the work reported in this
paper. The Titan V GPUs employed in this research were generously donated to NT
by NVIDIA. Matteo Ferrante and Tommaso Boccato are PhD students enrolled in the
National PhD in Artificial Intelligence, XXXVII cycle, course on Health and
life sciences, organized by Università Campus Bio-Medico in Rome (Italy). Marianna
Inglese is supported by the NANOINFORMATIX project which has received funding
from the European Union’s Horizon 2020 research and innovation programme under
grant agreement No 814426. The authors also acknowledge support from Imperial-NIHR-Biomedical-Research-Centre and Imperial-Experimental-Cancer-Medicine’s-Centre awards.References
1. Gao,
H. & Jiang, X. Progress on the diagnosis and evaluation of brain tumors. Cancer
Imaging 13, 466–481 (2013).
2. Dubash, S. R. et al. Clinical translation of
18F-fluoropivalate – a PET tracer for imaging short-chain fatty acid
metabolism: safety, biodistribution, and dosimetry in fed and fasted healthy
volunteers. Eur J Nucl Med Mol Imaging 47, 2549–2561 (2020).
3. Tofts, P. S. et al. Estimating kinetic parameters
from dynamic contrast-enhanced T(1)-weighted MRI of a diffusable tracer:
standardized quantities and symbols. J Magn Reson Imaging 10,
223–232 (1999).
4. Yankeelov, T. E., Rooney, W. D., Li, X. & Springer,
C. S. Variation of the relaxographic ‘shutter-speed’ for transcytolemmal water
exchange affects the CR bolus-tracking curve shape. Magn Reson Med 50,
1151–1169 (2003).
5. Li, X., Rooney, W. D. & Springer, C. S. A unified
magnetic resonance imaging pharmacokinetic theory: intravascular and
extracellular contrast reagents. Magn Reson Med 54, 1351–1359
(2005).
6. Inglese, M. et al. Reliability of dynamic
contrast-enhanced magnetic resonance imaging data in primary brain tumours: a
comparison of Tofts and shutter speed models. Neuroradiology 61,
1375–1386 (2019).
7. Inglese, M., Duggento, A., Boccato, T., Ferrante, M.
& Toschi, N. Spatiotemporal learning of dynamic positron emission
tomography data improves diagnostic accuracy in breast cancer. Annu Int Conf
IEEE Eng Med Biol Soc 2022, 186–189 (2022).
8. Parker, G. J. M. et al. Experimentally-derived
functional form for a population-averaged high-temporal-resolution arterial
input function for dynamic contrast-enhanced MRI. Magn Reson Med 56,
993–1000 (2006).