Stem cell transplants are in clinical trials for treating or slowing a myriad of diseases. Determining the location and number of transplanted cells, both directly and serially after delivery, will be crucial for monitoring treatment success. MRI-based single cell detection can enable this. To achieve single cell detection by MRI, cells are labeled with superparamagnetic iron oxide particles causing punctate hypointensities (spots) in T₂*-weighted MRI. Manual enumeration of these cells in 3D MRI is a tedious and inefficient task that is prone to subjectivity and inaccuracy. This is one key obstacle in quantifying in vivo cell number folloiwng transplant. Unfortunately, there is no commercial tool to automatically and accurately detect, locate and quantify transplanted cells. Therefore, we developed a machine learning based automatic strategy that intelligently quantifies cells in an MRI with an accuracy of 99.8% in vitro and 94.6% in vivo. To demonstrate accuracy and reliability, we evaluated our strategy under a diverse set of 34 testing conditions. Based on our detailed evaluation and its comparison with previous state-of-the-art, this is the most accurate cell quantification approach yet reported.

The workflow was to develop and evaluate the machine learning approach on well defined in vitro samples with known numbers of cells, then apply this to in vivo MRI.

MRI: Agarose phantoms were constructed with a dilute, known number of 4.5 micron diameter MPIOs (Dynal). Each bead has 10 pg of iron, simulating a labeled cell. We performed MRI of 29 agarose phantoms at 7T using a 3D FLASH sequence with TR/TE=30/10ms, and 100 μm isotropic voxels. We further performed in vivo MRI of 5 rat brains. 3 rats were injected intracardiac with MPIO-labeled MSCs, delivering cells to the brain – 2 rats were naïve. MRI was identical to in vitro.

Machine learning based cell quantification: Machine learning is category of artificial intelligence that enables a computer or a machine to learn a specific task from humans and then automatically perform that task without manual assistance. Recently, deep learning has proved to be the most successful machine learning algorithm. We use this algorithm to learn spot definitions and then automatically quantify them in unseen MRI. We explain our approach in a step-wise manner:

1. Since our approach learns from humans, therefore as a first step a human expert manually click on spots in MRI using our in-house developed software. These clicks will define ground-truth on spots in MRI.

2. The brain is segmented out of the MRI and then further over-segmented using superpixels[1] to finally create a set of patches (See Figure 1). Each patch may or may not contain a spot cross-section. Location of each patch is recorded and maintained.

3. Next, during the training/learning phase, the convolutional neural network deep learning algorithm evaluates patches that contain an expert’s ground-truth (spots) to the ones with no ground truth (non-spot patches) and learns a model containing spot definitions.

4. During the testing phase, a set of patches are extracted from a different unseen MRI and the learned model automatically detect spot patches in it.

5. Neighboring spot patches in consecutive slices are combined to form 3D spots.

Figure 2 shows the number of particles quantified in the 6 identical in vitro samples from data acquired with various imaging parameters. Theoretically, there should be 2400 spots detected based on phantom construction. The number of detected spots ranges from ~2000 to ~3000 with increasing number detected as the TE increases from 10 to 30. Comparing row D with A, and E with B, we see that our approach shows robustness to noise in MRI.

We used the area under the curve (AUC) as a standard measure to evaluate the performance of the trained model, taking the manual spot definitions as ground truth. Using a set of 4 in vitro scans, our approach achieved cell detection accuracy of 99.8%. We show the ROCs for in vitro tests in Figure 3.

In vivo MRI of rats injected with labeled MSCs (n=3) had dark spots distributed throughout the brain whereas control animal scans (n=2) did not show spots. Single cell distribution in the brain was verified by histology. This set of in vivo MRI scans was evaluated using the standard AUC measure, achieving 94.6% accuracy. Cell numbers quantified from the 3 in vivo brain data with injected cells, ranged from 1417 to 3900. Figures 4 and 5 show the automatically detected spots in an in vivo and in vitro slice of MRI respectively.These results are significantly superior to what state-of-the-art achieved on the same data[2].