By Laura Tamayo
A key technology for detecting cancer and heart diseases, are performed on 62 million patients annually in the United States. That���s close to one-fifth of the population, every year, going to hospitals, imaging centers, and laboratories to have a scan of their internal organs made into a three-dimensional image.
���The quality of reconstructed images directly affects how accurately clinicians can visualize anatomical structures and identify abnormalities,��� says Professor . ���This is an important problem in medical imaging.���
Last year, Benkarroum taught a senior named Angel Ostolaza at Borough of Manhattan Community College. Under Benkarroum���s mentorship, Ostolaza studied analytical reconstruction methods in medical imaging. He presented this research at the at Harvard.
���Professor Benkarroum���s project was the first time I did research,��� says Ostolaza. ���You start with something simple and then you build up into actually going through with the reconstructions and implementing your own algorithms. As I got more comfortable, I got to play around with my own ideas.���
Ostolaza was interested in improving the quality of (FBP). FBP is one of the most widely used scanning techniques, but it is very sensitive to noise (visual grain that blurs the image), incomplete sampling, and other limitations. Ostolaza combined that method with iterative reconstruction ��� which is more expensive and time-consuming, but also more resilient.
Ostolaza simulated raw CT data then introduced noise at different levels, from 0% to 50% blockage. He then utilized six different reconstruction methods to see how they dealt with the noise. Ostolaza measured the cost-quality tradeoff by weighing the image���s final quality and the time each process took.
Of the six methods, four used (ART), which computes the image twice: first as an approximation, then with more corrective details. Some involved Blank Seed, which means the initial image is blank. Regulation implies mediation and editing.
Ultimately, Ostolaza found that ART + Regulation (with either Blank or FBP Seed) yielded images closest to the simulated sinogram. However, at 25 iterations, ART takes around 1.7 times longer than FBP alone. At 50 iterations, it takes 3.4 times longer. The visual difference is strongest at higher noise levels (25���50%), where FBP struggles and ART maintains structural integrity. Thus, the greater computing cost and time spent is more justified at the higher noise levels.
Inspired by health researchers already using artificial intelligence, Ostolaza wants to combine machine learning with medical imaging. He is currently working on ideas for software that quickens the process of reconstruction while optimizing integrity.
More robust reconstruction algorithms will increase resilience to low-quality, noise-heavy data. This means medical professionals ��� even if their technology isn���t top-tier ��� will have an easier time reading the scans. Likewise, they can work quickly without sacrificing accuracy. In healthcare, time can mean the difference between life and death.
Research has shaped Ostolaza���s self-perception as both a student and a person in the world. In the fall, he���ll study at Baruch College. Then, he plans to go to graduate school.
���Research is the most important thing I���ve done as a student,��� says Ostolaza. ���When you���re going to the symposiums, you���re able to talk to people who are solving problems, obsessing over them in a way only people truly passionate about them can do.���
