AI in Medical Imaging

Non-Parametric Bayesian Deep-learning Methods for MRI Applications
Samah Khaled, PhD; Bosh (previously Technion)
Existing DNN methods for reconstruction and analysis often overlook uncertainty in their estimates or rely on strong assumptions about the prediction uncertainty, ultimately oversimplifying the complex underlying distribution. In our first two studies, we introduce a non-parametric Bayesian framework to estimate the uncertainty in DNN-based MRI tasks, employing stochastic gradient Langevin dynamics (SGLD). Implemented in MRI reconstruction and deformable registration, this method effectively characterizes the posterior distribution, enhancing uncertainty estimates related to out-of-distribution data, and produces more accurate predictions. Additionally, we present an end-to-end, uncertainty-aware slice-to-volume registration (SVR) model featuring a self-attention mechanism to eliminate uncertainty stemming from the noise in the input data. This model is designed for both real-time and retrospective correction of head motion in functional MRI (fMRI), instantly aligning acquired slices with a reference volume. Our research advances MRI analysis by effectively addressing uncertainty in DNN predictions, crucial for clinical reliability and accuracy.
Dr. Samah Khaled holds a Ph.D. from the Technion’s interdisciplinary program of applied mathematics, supervised by Dr. Moti Freiman. She currently serves as a Research Intern at Bosch Center for Artificial Intelligence. She previously earned her M.Sc. (Magna Cum Laude), under the supervision of Prof. Yehoshua Y. Zeevi, and B.Sc. degree, from the Technion’s ECE Department in 2017 and 2019, respectively. Her prizes and honors include the Ariane de Rothschild Women Doctoral fellowship, the Eric and Wendy Schmidt Postdoctoral Award for Women in Mathematical and Computing Sciences and Prof. Rahamimoff Travel Grant for Young Scientists, through the U.S.-Israel Binational Science Foundation (BSF).

Leveraging AI for Research: Enhancing Fetal Growth Restriction Prognostication with MRI
Aviad Rabinowich, MD, PhDc; Faculty of Medicine, TAU
Fetal growth restriction (FGR) is defined as a fetus that does not achieve its growth potential, complicating up to 10% of pregnancies. The diagnosis and risk stratification are typically done using ultrasound, which often lacks sensitivity for predicting perinatal adverse outcomes. Current measurement methods may be overly simplistic and fail to fully characterize the true deficits of the fetus. Our aim was to characterize pregnancies complicated by FGR. We employed several imaging techniques and deep learning/machine learning (DL/ML)-based tools to gain insights into the developing fetus. For instance, we demonstrated that FGR fetuses tend to have a leaner body habitus and that those classified as FGR with less fat are at a higher risk for adverse perinatal outcomes. To achieve this, we developed and utilized two in-house DL tools for subcutaneous fat tissue and fetal body segmentation.
Aviad Rabinowich, M.D., is currently a radiology resident at Tel Aviv Medical Center (Ichilov) and a Ph.D. candidate at Tel Aviv University. His main research focus is on improving the prognostication of fetal growth restriction (FGR or IUGR) pregnancies through MRI.

Enhanced Medical Image Segmentation Using Adapted and Fine-Tuned Segment Anything Model (SAM)
Tal Shaharabany, PhDc; Computer Science, TAU
The recently introduced Segment Anything Model (SAM) has demonstrated impressive image segmentation capabilities but its performance diminishes in Out-Of-Distribution (OOD) domains such as medical imaging. First, we introduce an encoder that replaces SAM's original conditioning, enabling automatic segmentation of medical images without further fine-tuning SAM. This encoder, trained via gradients from a frozen SAM, achieves state-of-the-art results across multiple medical image and video benchmarks. Second, we propose a minimally-guided zero-shot segmentation method leveraging SAM's ability to segment arbitrary objects in natural scenes. By extracting initial masks from self-similarity maps and applying test-time fine-tuning, we adapt SAM to the medical domain without requiring labeled medical data, using only a few foreground and background points on the test image. Experiments on diverse datasets validate the effectiveness of our approaches, with our code publicly available on GitHub.
Tal Shaharabany is a Ph.D. candidate at Tel Aviv University, specializing in computer vision and deep learning under the guidance of Prof. Lior Wolf. Their research focuses on advancing medical image segmentation techniques. Tal earned an M.Sc. in Electrical Engineering from Tel Aviv University, also under Prof. Wolf's mentorship. In addition to their research, Tal has been actively involved in teaching and assisting with courses on convolutional neural networks and deep learning at Tel Aviv University.