My research at MSKCC primarily focuses on hierarchical constrained optimization methods for Intensity Modulated Radiation Therapy (IMRT) treatment planning. The two main components of IMRT treatment planning system are a) Linear Accelator (LINAC) and b) Multi-Leaf Collimator (MLC). LINAC is responsible for producing high energy x-ray beam. MLC consists of many metal leaves that move to "block" or "pass" the x-ray beams that pass through it. The "blocking" or "passing" of x-ray beam by MLC modulates the intensity of x-ray beam. By virtue of this intensity modulation, x-ray can be focused more on tumor volume and less on normal organs. MLC is mounted on top of a gantry that rotates around a patient positioned on a couch. Figure 1 shows a Varian LINAC. An optimized dose distribution is a vital step of IMRT treatment to provide an efficient radiation therapy. The dose can be optimized by optimizing the x-ray fluence map. An efficient optimization algorithm can provide an optimized IMRT dose plan. Hierarchical Constrained Optimization or HCO is a widely used optimization technique that prioritizes the optimization objective functions and solve the problem in multiple steps. In IMRT HCO, objective function is prioritized based on the priority of the organ. For example, maximizing the x-ray dose at tumor or planned target volume (PTV) takes the highest priority followed by minimizing the dose at normal organs. At MSKCC, an in-house IMRT optimization routine named Expedited Constrained Hierarchical Optimization (ECHO) is developed that solves an optimization problem in 4 different steps. Step-1 maximizes the PTV coverage, step-2 minimizes the dose at organ-at-risk (OAR) that are located close to PTV and is at risk of getting higher dose and step-3 smoothes out the x-ray beam fluence map for better delivery efficiency. The final step also known as the "correction" step alleviates any discrepancies between optimized and final dose calculation.

Dose volume constraints (DVCs) play an important role in modern IMRT treatment. DVC allows only a certain percentage of volume of an organ to be overdosed and therefore can provide a clinically desirable dose distribution. DVCs are usually expressed as V (>d Gy) <v% which indicates that organ volume (V) that receives more than d Gy (Gray, unit of dose) is restricted to v% of volume V. Although, DVCs are crucial in controlling the amount of radiation dose in various organs, implementing DVC into IMRT optimization is a daunting task as it makes the optimization objective a non-convex function. DVC based optimization problem can be solved by a method named mixed integer programming (MIP). Although MIP can provide a ground truth solution, it is computationally expensive and takes many hours to solve. In my research, I adopt a "convex relaxation" method that simplifies MIP to provide an initial solution. I use this initial solution and apply an efficient heuristic to solve DVC based optimization problem. The proposed method is computationally faster and able to satisfy DVC as shown in figure 2.

My research at St Jude primarily focused on developing total-variation based noise reduction algorithms for low-dose cone-beam CT imaging. Cone-beam computed tomography (CBCT) has been extensively used in radiation therapy as a medical imaging technique to acquire a high-resolution volumetric image of a patient for treatment positioning. However, the repeated use of CBCT during the treatment course increases the risk of extra radiation dose delivered to patients. The extra radiation exposure to normal tissue during CBCT increases the risk of cancer and genetic defects. Therefore, the unwanted CBCT radiation dose must be minimized in order for the patients to truly benefit from the modern medical imaging techniques.

One way to reduce the CBCT dose is to acquire the CBCT projection data by reducing either the X-ray source tube current or the pulse duration. However, for lower current (mA) level or shorter pulse duration (ms), the projection image is contaminated with excessive quantum noise.

I was interested in reducing the noise level on the projected image that is acquired by low mAs CBCT protocols. A total-variation (TV) based noise reduction algorithm was studied and applied to a computer–simulated phantom, physical phantom and patient data. The algorithm had shown to have the potential in reducing the noise level for low-dose CBCT images without compromising the contrast and resolution of the images. Figure-3 shows the denoised images using TV-based algorithm.

My Ph.D. research mainly focused on simulation of microwave-induced thermo-acoustic tomography, abbreviated as MI-TAT, for breast and prostate cancer imaging. MI-TAT is an emerging medical imaging modality which combines both microwave imaging and ultrasound imaging for providing high-contrast and high-resolution images of cancerous tissues. In MI-TAT, biological cells are irradiated by short-pulsed microwave energy in the frequency range of 434 MHz-3 GHz. Absorption of this energy by tissues causes a thermo-elastic expansion of the cells and produces an acoustic/pressure wave. Detection of this acoustic wave by ultrasonic detectors can give crucial information about the dielectric properties (e.g. relative permittivity and electrical conductivity) of the tissues. It has been well proved that malignant tissues usually exhibit different electrical conductivity from normal tissues because of different concentrations of ions and water and absorb more microwave energy. MI-TAT exploits this different microwave energy absorption characteristic to give an imaging contrast between malignant and normal tissues.

I developed a finite element method (FEM) based forward and inverse modeling of thermo-acoustic tomography to reconstruct conductivity distribution in a computer-simulated breast and prostate like geometry. The algorithm was written in C++/MATLAB. Since cancerous tissues possess different conductivity from normal tissues, reconstructing conductivity distribution using MI-TAT can give a precise location of cancerous tissues. Figure-4 shows a simulated reconstructed image using TAT for a breast-like geometry.