Machine Learning Powered Automatic Organ Classification for Patient Specific Organ Dose Estimation Junghwan Cho, Eunmi Lee, Hyunkwang Lee, Bob Liu, Xinhua Li, Shahein Tajmir, Dushyant Sahani, and Synho Do* Lab. Of Medical Imaging and Computation Massachusetts General Hospital and Harvard Medical School MGH and BWH CCDS
Disclosure CRICO: grant T-PLUS: sponsored research agreement McCoy Medical Technologies: Non-commercial research license NVIDIA: GPU system donation
Contents How to estimate radiation dose? How to use Organ Recognition Machine Learning algorithm for Organ Dose Estimation? Results/Comparison Conclusions
Radiation Dose Estimation CTDI (CT Dose Index) : Average dose to the central slice of a cylindrical PMMA phantom CTDI w, CTDI vol CTDI vol to organ dose conversion coefficients Patient specific organ dose estimation with AI
CTDI w and CTDI vol CTDI varies across the field of view To represent dose for a specific scan protocol #SIIM17 CT Dose Estimation by John Boone
DLP (Dose Length Product) and E (Effective Dose) - To better represent the overall energy delivered - To reflect the total energy absorbed - To consider biological sensitivity of the tissue or organ - Single dose parameter that reflects the risk of a nonuniform exposure for an equivalant whole body EE(mmmmmm) kk DDDDDD
Approaches Accurate Fast Unrealistic Measurements Simulations General Slow Computational Challenge
Our Approach Organ Recognition AI ImPACT Simulation Patient Specific Scanner Model Scan/Acquisition Parameters Look-up table mapping
Proof of Concept Training Data (A) Whole-body CT images (B) Six body parts
How many data set? Test data set Learning Curve
Body Part Recognition AI (6 Classes) Organ Detection AI (16 Classes) ImPACT CT Dosimetry Slabs
HEAD1: Brain HEAD2: Eye Lens HEAD3: Nose HEAD4: Salivary Gland HEAD5 Body Part Classification (16 classes) #SIIM17 TRUNK1: Upper Lung TRUNK2: Thymus TRUNK3: Heart TRUNK4: Chest (Liver Yes) TRUNK5: Abdomen 1 TRUNK6: Abdomen 2 TRUNK7: Pelvis 1 TRUNK8: Pelvis 2 TRUNK9: Urinary Bladde LEG 1 LEG 2 SIEMENS Biograph 64 Scanner CT Nk/Ch/Abd I+ 2.0 B31f PETCT 2 mm slice thickness 100 male and 100 female patients
Organ Recognition AIs Training a 22 layers deep convolutional neural network ( Male and Female) Male: Train-> 70 Patients, Total: 20, 424 image slices Test -> 30 Patients, Total: 8,564 image slices Training data sets of male Female: Train-> 70 Patients, Total: 19,335 image slices Test -> 30 Patients, Total: 8,017 image slices
Test result of body part classification with 30 patients each gender #SIIM17 Male ACC=98.94%, AUC=0.999, AP=0.998 Female ACC=99.11%, AUC=1.000, AP=0.998
Software Flowchart DICOM Read (Patient) Machine Learning 1 (Body Part Classification) Scanner & Scan Parameter ImPACT Dose Calculator Scan Region ImPACT based Organ Dose Estimation DICOM to Images Conversion Measurement of Patient Effective Diameter Correction Factor (Regression Model) 1.8 2 Heart 1.6 Machine Learning 2 (Organ Segmentation) MGH PODE (Coase Tuning to Patient Weight) Organ Detection & Segmentation + MGH PODE (Fine Tuning to Patient Organ) Normalized Organ Dose by CTDIvol 1.4 1.2 1 0.8 0.6 200 250 300 350 400 450 Patient Effective Dia(mm)
Patient Specific Organ Reg. AI Each Patient : Unique Organ Shape and Size Patient Organ Detection and Segmentation : Organ Volume Estimation (Lung, Liver, Kidney, Urinary Bladder, Muscles etc) Our Solution Machine Learning Powered Automatic Organ Classification. Liver Urinary Bladder Liver Urinary Bladder #SIIM17 Same BMI
Scan Region Determination by Body Part Classification (Chest Scan) #SIIM17 Coronal view Axial slices Classification results Corresponding scan range of ImPACT Slice #1 Slice #66 Slice #134 Slab #170 Slab #83
Scan Region Determination by Body Part Classification (Abdomen-Pelvis Scan) #SIIM17 Coronal view Axial slices Classification results Corresponding scan range of ImPACT Slice #1 Slice Slice #32 #66 Slab #126 Slice #66 Slab #11
Radimetrics vs MGH Personalized Organ Does Estimator (PODE) Phantom Model SSDE (Size Specific Organ Dose Estimate) Organ Specific (Shape & Volume) Radimetrics* Mathematical Model (20 phantom models [6/3/11]) Yes (Patient Effective Diameter) No MGH PODE Mathematical Model (ImPACT 1.0.4 Patient Dosimetry) Yes (Patient Effective Diameter) Yes * Radiation Dose Calculations in Radimetrics Enterprise Platform by Bayer (RAD-INF-13-05968, March 2014) #SIIM17
Organ Dose Estimation Comparison (1/3) #SIIM17 Patient Weight=75kg Patient Weight=86kg Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE 0 1 2 3 4 5 6 7 8 9 Organ Dose [msv] 0 2 4 6 8 10 12 Organ Dose [msv]
Organ Dose Estimation Comparison (2/3) #SIIM17 Patient Weight=93kg Patient Weight=104kg Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE 0 5 10 15 Organ Dose [msv] 0 2 4 6 8 10 12 14 16 18 Organ Dose [msv]
Organ Dose Estimation Comparison (3/3) #SIIM17 Patient Weight=113kg Patient Weight=122kg Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE Brain Eye Lenses Salivary Glands Esophagus Thyroid Lungs Thymus Heart Liver Spleen Stomach Gall Bladder Pancreas Adrenals Kidneys Colon Small Intestine Uninary Bladder Testicles Skin Muscle Red Marrow Skeleton Impact 1.04 Radimetrics MGH PODE 0 5 10 15 20 Organ Dose [msv] 0 2 4 6 8 10 12 14 16 18 Organ Dose [msv]
Conclusions Body part/organ recognition is essential in automatic medical image analysis : anatomy identification and organ segmentation Compared to text-based body part retrieval in DICOM header, image based organ recognition is accurate and fast with machine learning (ML) algorithm. We implemented a ML powered automatic organ classifier for CT datasets with a deep convolutional neural network (CNN) followed by an organ dose calculation. We labeled 16 different organs from axial views of CT images. The resultant classified organ was automatically mapped to the slab number of a mathematical phantom to determine the scan range of ImPACT CT dose calculator. This technique can be used for patient-specific organ dose estimation since the location and sizes of organs for each patient can be calculated independently.