Medical Physics Works: Machine Learning in Radiation Oncology

Medical Physics Works: Machine Learning in Radiation Oncology

Automated Brain Tumors Segmentation and Patient Overall Survival Predictions with Machine Learning  Tumor  segmentation  (delineating...

Medical Physics Works: Machine Learning in Radiation Oncology

Medical Physics Works: Machine Learning in Radiation Oncology: Automated Brain Tumors Segmentation and Patient Overall Survival Predictions with Machine Learning  Tumor  segmentation  (delineating...

Medical Physics Works: Intesity Modulated Radiation Therapy Quality Assur...

Medical Physics Works: Intesity Modulated Radiation Therapy Quality Assur...: Patent-Specific  IMRT  QA: ِِA Developed Program/Software for Dose Distributions Evaluation (Gamma Calculation) The work is about the do...

Machine Learning in Radiation Oncology

Automated Brain Tumors Segmentation and Patient Overall Survival Predictions with Machine Learning 

Tumor segmentation (delineating a line around the tumor, or contouring) is conventionally used to be done manually by the radiation oncologist which requite considerable effort and time consuming. Auto-process of this task is significantly reducing the 

challenges in developing an automated tumor delineation tool. 

associated with quality and safety of the patient treatment plan, 

Accurately drawing the tumor in the treatment plan on patient CT is one of the most big challenge in patient treatment chain in radiation oncology. Up to date, there is no software that is quite good enough for this task. In this work, we aimed to develop an automated tool for brain segmentation on a multi modal MRI scans based on state-of-the-art machine learning algorithms and predicts the patient overall survival.   

Fig. 1.Top, Predicted segmentation labels (green), edema tumor, on a T2-Flair MRI BraTs'2017 validation data (slice# 84:2:98) with our automated model. Bottom, A tumor on MRI (left) and its predicted segmentation (right) with coronal, sagital, and axial view. 

Fig. 2. The plot of the OS classification predicted regions (Left) and the confusion matrix (Right). On the confusion matrix, the rows correspond to the predicted class (Output Class), and the columns show the true class (Target Class). The diagonal cells show for how many (and what percentage) of the examples the trained algorithm correctly estimates the classes of observations. That is, it shows what percentage of the true and predicted classes match. The off diagonal cells show where the classifier has made mistakes. The column on the far right of the plot shows the accuracy for each predicted class, while the row at the bottom of the plot shows the accuracy for each true class. The cell in the bottom right of the plot shows the overall accuracy. 

Overview about machine learning

What is Machine learning?

Machine Learning (ML) is actually a set of rules that a computer develops on its own to correctly solve problems. The basic idea is that a ML computer will find patterns in data and then predict the outcome of something it has never seen before. ML is a critical component to any Artificial Intelligence (AI) development.

Machine learning is classified into three categorizes: Supervised Learning, Reinforcement Learning (semi-supervised), and Deep Learning (unsupervised). 

a) Supervised Learning which is a type of machine learning that feeds a computer system many (thousands, millions or even billions) of examples of a given item and having the computer calculate the similarities between those items so that it can recognize other examples of that item which it has not seen yet. 

b) Reinforcement Learning which is a type of machine learning that tells a computer if it has made the correct decision or the wrong decision. With enough iterations a reinforcement learning system will eventually be able to predict the correct outcomes and therefore make the ‘right decision’.

c) Deep Learning is a type of machine learning that requires computer systems to iteratively perform calculations to determine patterns by itself. 

Why ML pop-up recently?
Until recently ML was not possible because we lacked the very large data sets computers require to find patterns in, the storage capacity to keep all of that data and the computing power to find those patterns in a reasonable amount of time. All three of these factors have now changed: 

1. With the advent of parallel processors, usually GPU’s (Graphics Processing Unit’s) the computing power issue has been advanced to the point where Machine Learning is now possible.
2. ‘Big Data’ has now collected enough data samples for Machine Learning to work with.  

Intesity Modulated Radiation Therapy Quality Assurance

Patent-Specific IMRT QA: ِِA Developed Program/Software for Dose Distributions Evaluation (Gamma Calculation)

The work is about the dose evaluation of two dose distributions (reference and evaluated distributions, or measurement and plan dose distributions). We developed a program does the calculation of dose difference (DD), distance to agreement (DTA) and gamma index. The program is validated with Daniel Low results for two synthesized relative dose distributions that published in AAPM TG-120 report. 

Fig. 1: Validation our program with Daniel Low works (AAPM TG-120). Left, Daniel Low works: a) reference distribution, b) evaluated distribution, c) dose difference (b) – (a), d) distance to agreement (DTA), e) composite (gamma binary), and f) gamma index (3% 3 mm)., and Right, Reproducing Daniel low works for verification (our program results).

The program is also validated with I'mRT MATRIX software for one clinical case (prostate) and the amma index results are compared.

Fig. 2: Validation the program with ImRT MATRIX software (IBA Dosimetry) for a clinical case (prostate).  Left, Gamma index for a prostate case (ImRT MATRIX software).  Right, Gamma index for a prostate case (our program). 

Patient-Specific IMRT/VMAT QA: A Developed Program/Software for Patient-Specific IMRT/VMAT QA using Log File Analysis (Calculations-based Method)

Current QA is performed as 1) Pre-treatment verification of fields (Checks of leaf motion, Transfer of data, & General deliverability of plan), 2) Second check software, and /or 3) Frequent imaging.

What is missed? Verification of correct delivery when the plan is delivered to the patient on a daily basis. 

The American College of Radiology (ACR) recently (August 2015) published a statement on their website answering the question of whether the ACR Radiation Oncology Practice Accreditation (ROPA) program considers log files generated from the treatment machine an acceptable “alternative” measurement for IMRT/VMAT for patient QA. The ACR states that log files are acceptable as long as they are generated using the patient’s IMRT/VMAT plan before the start of patient treatment, and that the medical physicist, along with the radiation oncologist have assured that the measurements verify the actual radiation doses the patient will receive. Two-dimensional detector arrays are specified as another ‘alternative’ method in the ACR-ASTRO Practice Guidelines for IMRT.

We have developed a software to analyze the MLCs performance during the IMRT delivery.

Fig. 1. IMRT QA using DynaLog file analysis method. a) leaf position error (mm), b) leaf gap error (mm), c) velocity error (cm/s), d) photon fluence error (%), and e) leaf deviation value (mm) (root mean square of leaf position).

Fig. 2: A hand-made DMLCs sequence.

Monte-Carlo Simulation

MATLAB-Based Monte-Carlo Calculations Algorithm for Radiation Transport in Medium for Educational Purposes

We have written a code that simulates the transport of photons in a medium. Photons of 50 keV energy are launched in the z-direction towards a phantom as a beam of 8 cm x 8 cm field size with 100 cm SSD. Photon distance to next interaction, interaction type (Photoelectric effect, Compton scattering and Rayleigh scattering), and scattered photon directions (CO & RA) are sampled using the cross section data and random numbers generator.

Fig. 1: Left: Schematic diagram showing the geometry. Right: An image of the dose deposited in the surface plane of the phantom, middle plane, exit plane, and the dose deposited in the z-direction in the center of the beam (y-view).

Fig. 2: Photons interaction and dose deposition during the simulation.