RADIOMICS: potential role in the clinics and challenges Dr. - - PowerPoint PPT Presentation

RADIOMICS: potential role in the clinics and challenges

• Dr. Francesca Botta

Medical Physicist Istituto Europeo di Oncologia (Milano)

27 giugno 2018 Dipartimento di Fisica – Università degli Studi di Milano

RADIOMICS: definition

Radiomics is a field of medical study that aims to extract large amount of quantitative features from medical images using mathematical algorithms. These features, termed radiomic features, have the potential to uncover disease characteristics that fail to be appreciated by the naked eye. The hypothesis of radiomics is that the distinctive imaging features between disease forms may be useful for predicting prognosis and therapeutic response for various conditions, thus providing valuable information for personalized therapy. Radiomics emerged from the medical field of oncology and is the most advanced in applications within that field. However, the technique can be applied to any medical study where a disease or a condition can be tomographically imaged.

RADIOMICS: definition & workflow

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

Predictive / Prognostic models

RADIOMICS: history

RADIOMICS: history

Computational power

Haralick, 1973:

Big data analysis

…extraction of LARGE amount

• f quantitative features

Clinical data availability

Store & retrieval of large amount of clinical data and images Digital Imaging

Texture analysis

Extraction of quantitative parameters from images

• omics

Experience from other fields (Molecular biology, genetics, …)

RADIOMICS: history

RADIOMICS: history

RADIOMICS: potential role in the clinics

1. Imaging is routinely performed for oncologic patients:

• diagnosis
• treatment planning
• follow up

2. Imaging is not invasive and minimally detrimental invasive alternatives: biopsy, blood sampling 3. Radiomics quantifies the properties of the whole volume reduced risk of under-sampling as compared to e.g. biospy

• plenty of retrospective data
• database continuously updated
• no additional cost

Considering that:

• no additional patient discomfort
• more complete information

RADIOMICS: history

20 40 60 80 100 120 1985 1986 1987 1995 1998 2000 2001 2002 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018

Pubmed search

Radiomics

• Differentiate malignant / benign tissue
• Tumour staging: differentiate between early and advanced stage disease
• Prognostic models: correlation with survival
• Predictive models: predict treatment response (chemotherapy, radiation therapy)
• Assessment of the metastatic potential of tumours
• Assessment of cancer genetics / biological or histopathological properties

(biological basis of clinical application of radiomics)

• Improve predictivity of models based on clinical,biological, genetic data

RADIOMICS: potential role in the clinics

RADIOMICS: potential role in the clinics

RADIOMICS: potential role in the clinics

RADIOMICS: challenges

Models generalizability

RADIOMICS: challenges

RADIOMICS: challenges

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

Data quality: «imaging biomarkers» are needed

RADIOMICS: challenges

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

Imaging biomarker: Which requirements?

RADIOMICS: challenges

Interpretation?

RADIOMICS: workflow

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

• 1. Image acquisition
• CT images:

the voxel intensity describes the composition and the density of the tissue

• PET images:

the voxel intensity is a measure of the concentration of the radiotracer

• MR images:

according to the sequence applied, the voxel intensity can be representative of different properties of the tissue (relaxation times T1, T2, proton density), diffusion, perfusion, …

RADIOMICS: workflow

Discrete sampling

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

RADIOMICS: workflow

The Volume Of Interest is a 3D array of numbers, from which many different parameters can be calculated

VOI: Volume Of Interest

• 2. Region segmentation
• Manual segmentation
• Semi-automatic / Automatic segmentation algorithms
• Machine learning

RADIOMICS: workflow

RADIOMICS: workflow

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

RADIOMICS: features extraction

Sha Shape fea eature res: describe the shape of the Region Of Interest in 3D

Geometric properties, like the Volume, the maximum diameter or the 3 diameters along the 3 orthogonal directions, the maximum surface

RADIOMICS: features categories

Histogram-based (First order statistics) features:

Describe the distribution of values of individual voxels without concern for spatial relationships Different spatial arrangement BUT Same Histogram!

RADIOMICS: features categories

His istogra ram-based (Fi (Firs rst ord

• rder statistic

ics) ) fea eature res:

Describe the distribution of values of individual voxels without concern for spatial relationships (histogram-based methods as mean, median, maximum, minimum, uniformity or randomness (entropy) of the intensities, skewness (asymmetry) and kurtosis (flatness) of the histogram of values.

RADIOMICS: features categories

Textural (Second order statistics) features:

“Textural” features describing statistical interrelationships between voxels with similar (or dissimilar) values and take into account the spatial arrangement of the values. « Haralick features»

GLCM: Gray Level Cooccurrence Matrix Image ….

RADIOMICS: features categories

v

Image GLRLM: Gray Level Run Length Matrix ….

RADIOMICS: features categories

Higher order statistics features:

Higher-order statistical methods applies filter grids or mathematical transforms to the image (for example, to extract repetitive or nonrepetitive patterns)

• wavelet transform;
• fractal analyses, wherein patterns are imposed on the image and the

number of grid elements containing voxels of a specified value is computed;

• Laplacian transforms of Gaussian bandpass filters that can extract areas

with increasingly coarse texture patterns from the image; Radiomic features calculation is performed on the filtered or decomposed images in order to extract multiple or more informative parameters from a single image.

Fractal dimension

RADIOMICS: features categories

RADIOMICS: challenges

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

The voxel intensity is the starting point for features calculation

RADIOMICS: challenges

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

• A. What affects the voxel intensity?

Does it also affect the radiomic feature value?

RADIOMICS: challenges

Discrete sampling

Partial Volume Effect

• 1. Image acquisition affects the informative content of the image

RADIOMICS: challenges

• 1. Image acquisition affects the informative content of the image

Voxel size: Pixel size Slice Thickness

Discrete sampling

RADIOMICS: challenges

Spatial resolution

Image from Soret et al., ‘‘ Partial Volume Effect in PET tumour imaging’’, Journal of Nuclear Medicine (2007), 48(6): 932-945

Scanner properties Acquisition protocol Reconstruction algorithm

• 1. Image acquisition affects the informative content of the image

RADIOMICS: challenges

Noise Scanner properties Acquisition protocol Reconstruction algorithm

• 1. Image acquisition affects the informative content of the image

RADIOMICS: challenges

Bit depth

• 1. Image acquisition affects the informative content of the image

RADIOMICS: challenges

• 2. Image post-processing

Post-reconstruction image filtering Useful for physicians, visual inspection, clinical reporting …impact on quantification? Discretization N possible values in the image Size of GLCM: NxN

RADIOMICS: challenges

• A. …what affects the voxel intensity?

Does it also affect the radiomic feature value?

Pixel size Slice Thickness Scanner properties Acquisition protocol Reconstruction algorithm Bit depth Post-reconstruction image filtering Discretization Reproducibility: different modalities are comparable? Repeatability: one modality more stable than others?

RADIOMICS: challenges

• A. …what affects the voxel intensity?

Does it also affect the radiomic feature value?

Pixel size Slice Thickness Scanner properties Acquisition protocol Reconstruction algorithm Bit depth Post-reconstruction image filtering Discretization Reproducibility: different modalities are comparable? Repeatability: one modality more stable than others?

RADIOMICS: challenges

• A. …what affects the voxel intensity?

Does it also affect the radiomic feature value?

Pixel size Slice Thickness Scanner properties Acquisition protocol Reconstruction algorithm Bit depth Post-reconstruction image filtering Discretization Reproducibility: different modalities are comparable? Repeatability: one modality more stable than others?

Repeatability - CT

RADIOMICS: challenges

Concordance Correlation Coefficient > 0.9

Feature

Acquisition/ Reconstruction setting

Example: Phantom experiment, CT % Coefficient of Variation among 10 consecutive acquisitions

Repeatability - CT

RADIOMICS: challenges

Different acquisition/reconstruction settings have different repeatability

RADIOMICS: challenges

Repeatability - PET

RADIOMICS: challenges

…about Repeatability:

It should be assessed according to the Dynamic Range observed in-vivo, and according to the difference between the groups Distribution of values assessed after repeated acquisitions Feature value

RADIOMICS: challenges

…about Repeatability:

It should be assessed according to the Dynamic Range observed in-vivo, and according to the difference between the groups Distribution among many subjects

• f a group

Feature value

RADIOMICS: challenges

…about Repeatability:

It should be assessed according to the Dynamic Range observed in-vivo, and according to the difference between the groups Feature value Group 1 Group 2

Reproducibility - CT

RADIOMICS: challenges

Feature

Acquisition/ Reconstruction setting

Example: Phantom experiment, CT Feature Values obtained for 40 different acquisition and reconstruction settings

Reproducibility - CT

RADIOMICS: challenges

Some features are highly dependent on acquisition/reconstruction settings

RADIOMICS: challenges

Image processing

RADIOMICS: challenges

Computed Tomography – CT Positron Emission Tomography – PET Magnetic Resonance - MR

• A. What affects the voxel intensity?

Does it also affect the radiomic feature value?

• B. Other factors affecting the radiomic feature value?

RADIOMICS: challenges

• 3. Volume: segmentation

RADIOMICS: challenges

• 3. Volume: size

Small volumes, not only partial volume effect:

RADIOMICS: challenges

Rigorous methodology Standardization

RADIOMICS: challenges

Rigorous methodology Standardization Ask yourself:

is the feature really quantifying a biological property? Or… am I «measuring» noise? am I «measuring» volume?

RADIOMICS: challenges

Rigorous methodology Standardization Ask yourself:

• A. is the feature really quantifying

a biological property? Or… am I «measuring» noise? am I «measuring» volume?

• B. can I explain the meaning of the

features extracted from images?

Redundancy:

High correlation among features Hierarchical Clustering High intra-cluster correlation Low inter-cluster correlation

RADIOMICS: statistical analysis

RADIOMICS: the future…

Machine Learning

Data analysis

• radiomic features
• clinical data

Feature calculation

• image + Region Of Interest
• clinical data

Segmentation

• image
• clinical data