RADIOMICS AND
MACHINE LEARNING
METHODS FOR
COMPUTER-AIDED
PROSTATE CANCER
ASSESSMENT
•Author: Pablo Laso-Mielgo
•Supervisor: Norberto Malpica
•Co-supervisor: Sergio Dominguez
Contents
•Prostate cancer MRI and Previous
Studies.
•Objectives.
•Methodology.
•Results.
•Conclusions.
•Future Implementations.
Prostate cancer
MRI and
Previous
Studies
Prostate cancer MRI
and Previous Studies
•+27.6% PCa new cases by 2040.
Prostate cancer MRI
and Previous Studies
•+27.6% PCa new cases by 2040.
•Different CAD systems (lung nodules,
breast cancer, colon cancer, …).
Prostate cancer MRI
and Previous Studies
•+27.6% PCa new cases by 2040.
•Different CAD systems (lung nodules,
breast cancer, colon cancer, …).
•T2, DWI, and ADC maps.
Prostate cancer MRI
and Previous Studies
•+27.6% PCa new cases by 2040.
•Different CAD systems (lung nodules, breast
cancer, colon cancer, …).
•T2, DWI, and ADC maps.
•Previous studies on PCa Radiomics suggest:
•Low T2 efficiency.
•mp-MRI (using different image modalities).
•Feature selection methods (mRMR).
•SVM, RF, and LR.
Objectives
•Development of a Computer tool for
Prostate Cancer assessment (CS/CI).
•Study complete PI-RADS (1-5)
classification feasibility.
•Assess the efficiency of radiomics-driven
algorithms.
•Find most relevant radiomics features.
Methodology
•Image Acquisition and
Processing
•Image Acquisition and Processing.
•Segmentation
•Registration
•Feature Extraction.
•Data Analysis and Processing.
•Feature Selection.
•Classifiers/models.
Methodology: Image Acquisition
and Processing
•T2, DWI (1000s/mm2).
•DWI (b=1000) + DWI (b=0) →ADC
maps.
•DICOM →NIfTI.
Methodology: Image Acquisition
and Processing
•T2, DWI (1000s/mm2).
•DWI (b=1000) + DWI (b=0) →ADC
maps.
•DICOM →NIfTI.
Methodology: Image Acquisition
and Processing
•T2, DWI (1000s/mm2).
•DWI (b=1000) + DWI (b=0) →ADC
maps.
•DICOM →NIfTI.
Methodology: Segmentation
•Manual Segmentation (>100 * 3).
•Training Radiomics models.
•Training CNNs *.
•Automatic Segmentation *.
* MSc thesis (Sergio Domínguez)
Methodology: Registration
•T2 → DWI/ADC mapping.
•Python (affine matrix) or 3DSlicer.
•Joint findings for both modalities (if same
lesion).
ITK-Snap
Python
Methodology: Feature Extraction
•Python.
•pyRadiomics.
•Inputs:
•.yaml file →specify features.
•Image.
•Mask.
•Output:
•Feature matrix.
Methodology: Data Analysis and Processing
Methodology: Data Analysis and Processing
Methodology: Data Analysis and Processing
Methodology:
Data Analysis and Processing
•Correlation matrix.
•Highly correlated features.
Methodology: Feature Selection
•Filters
•Chi-squared.
•ANOVA.
•Pearson’s coeff.
•Mutual Information.
•Wrappers
•LR, DT, RF, SVM, KNN, Bagging, and GBM.
•Embedded
•Lasso.
•Ridge.
•Dimensionality reduction
•PCA.
ADC feature selection
T2 feature selection
Methodology: Classifiers
•SVM.
•LR.
•RF.
•Ridge.
•KNN.
•Bagging.
•GBM.
ADC feature selection
T2 feature selection
Results
•Modality-specific →CS/CI
•mp-MRI (t2w+DWI+ADC) •CS/CI
•PI-RADS (1-5)
CS/CI classification
using DWI only
CS/CI classification
using DWI only
DWI
CS/CI classification
using T2 only
CS/CI classification
using ADC only
CS/CI classification
using mp-MRI
CS/CI classification
using mp-MRI
CS/CI classification
using mp-MRI
PI-RADS (1-5)
classification
using mp-MRI
Conclusions
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
•ADC maps: higher accuracy after gamma correction.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
•ADC maps: higher accuracy after gamma correction.
•Top-ranked features: Gray values, Entropy, and Energy.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
•ADC maps: higher accuracy after gamma correction.
•Top-ranked features: Gray values, Entropy, and Energy.
•New pyradiomics matrices: GLCM along with GLRLM, GLDM, NGTDM, GLSZM.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
•ADC maps: higher accuracy after gamma correction.
•Top-ranked features: Gray values, Entropy, and Energy.
•New pyradiomics matrices: GLCM along with GLRLM, GLDM, NGTDM, GLSZM.
•ML models: not only limited to SVM.
Conclusions
•Complete radiomics pipeline for prostate cancer assessment in MRI studies.
•High scores for images from different MRI scanners.
•DW images: most efficient in both mp-MRI and non-mp-MRI.
•ADC maps: higher accuracy after gamma correction.
•Top-ranked features: Gray values, Entropy, and Energy.
•New pyradiomics matrices: GLCM along with GLRLM, GLDM, NGTDM, GLSZM.
•ML models: not only limited to SVM.
•Complete PI-RADS classification: lower efficiency. Ambiguous (especially 3vs4-PIRADS).
Future Implementations
Future
Implementations
•1400-b-valued DW images.
•
•
•
Future
Implementations
•1400-b-valued DW images.
•3D modalities, such as LAVA or DISCO.
•.
•
Future
Implementations
•1400-b-valued DW images.
•3D modalities, such as LAVA or DISCO.
•Professional segmentation and labelling.
•
Future
Implementations
•1400-b-valued DW images.
•3D modalities, such as LAVA or DISCO.
•Professional segmentation and labelling.
•Larger dataset (especially for a mp-MRI, 1-5
PI-RADS classification).
Thank you!
Author: Pablo Laso-Mielgo.
Tutor: Norberto Malpica-Gonzalez.
Cotutor: Sergio Domínguez-Rodríguez.
