RATS: Rapid Automatic Tissue Segmentation in rodent brain MRI

Abstract

Background: High-field MRI is a popular technique for the study of rodent brains. These datasets, while similar to human brain MRI in many aspects, present unique image processing challenges. We address a very common preprocessing step, skull-stripping, which refers to the segmentation of the brain tissue from the image for further processing. While several methods exist for addressing this problem, they are computationally expensive and often require interactive post-processing by an expert to clean up poorly segmented areas. This further increases total processing time per subject.

New method: We propose a novel algorithm, based on grayscale mathematical morphology and LOGISMOS-based graph segmentation, which is rapid, robust and highly accurate.

Results: Comparative results obtained on two challenging in vivo datasets, consisting of 22 T1-weighted rat brain images and 10 T2-weighted mouse brain images illustrate the robustness and excellent performance of the proposed algorithm, in a fraction of the computational time needed by existing algorithms.

Comparison with existing methods: In comparison to current state-of-the-art methods, our approach achieved average Dice similarity coefficient of 0.92 ± 0.02 and average Hausdorff distance of 13.6 ± 5.2 voxels (vs. 0.85 ± 0.20, p<0.05 and 42.6 ± 22.9, p << 0.001) for the rat dataset, and 0.96 ± 0.01 and average Hausdorff distance of 21.6 ± 12.7 voxels (vs. 0.93 ± 0.01, p <<0.001 and 33.7 ± 3.5, p <<0.001) for the mouse dataset. The proposed algorithm took approximately 90s per subject, compared to 10-20 min for the neural-network based method and 30-90 min for the atlas-based method.

Conclusions: RATS is a robust and computationally efficient method for accurate rodent brain skull-stripping even in challenging data.

Keywords: Brain segmentation; Rat model; Small animal MR.

Figure 1

RATS consists of a series of grayscale (blue) and binary (purple) morphological operators, which create an initial surface for the graph-based segmentation algorithm.

Figure 2

Segmentation performance for RATS and atlas-based tissue classification on the T1-weighted in vivo rat dataset. The values reported are the average (standard deviation). “RATS > Atlas” and “RATS > PCNN” should be interpreted as better, i.e. higher Dice and Jaccard indices and lower maximum Hausdorff distance and computation time. RATS performs significantly better than both the other methods, for all measured criteria.

Figure 3

Segmentation performance for RATS and atlas-based tissue classification on the T2-weighted in vivo mouse dataset. The values reported are the average (standard deviation). “RATS > Atlas” and “RATS > PCNN” should be interpreted as better, i.e. higher Dice and Jaccard indices and lower maximum Hausdorff distance and computation time. RATS performs significantly better than both the other methods, for all measured criteria.

Figure 4

Typical segmentation comparison for the T1-weighted in vivo rat dataset: green: ground truth, yellow: RATS, red: atlas-based, pink: PCNN. The boundaries are made thicker for visibility. Top: best case, middle: typical case, bottom: worst case. Note that PCNN is missing the olfactory bulb and surrounding areas even in the best case. The atlas-based approach is often missing a significant chunk of the brainstem. RATS achieves satisfactory results even in the worst case, with minor defects near the olfactory bulb.

Figure 5

3D brain segmentation on the worst-case subject for the T1-weighted in vivo rat dataset: pink: PCNN, red: atlas-based, yellow: RATS, green: ground truth. PCNN is entirely missing the olfactory bulb and surrounding frontal regions. Additionally, both PCNN and the atlas-based approach are missing significant chunks of the brainstem. RATS produces near-perfect segmentation even for this challenging image. The “worst” subject was chosen as the subject that had the lowest Dice score average over the three methods. The cases where the atlas method had extremely poor performance due to registration failure were excluded from this consideration.

Figure 6

Typical segmentation comparison for the T2-weighted in vivo mouse dataset: green: ground truth, yellow: RATS, red: atlas-based, pink: PCNN. The boundaries are made thicker for visibility. Top: best case, middle: typical case, bottom: worst case. All three methods produce decent results; however, the brainstem region is observed to have very low signal, causing problems for all methods, to varying degrees. Additionally, the atlas-based method is observed to suffer from inaccurate registration. RATS produces near-perfect results with the exception of the brainstem region and reaches significantly more accurate segmentations compared to the other two methods.

Figure 7

3D brain segmentation on the worst-case subject for the T2-weighted in vivo mouse dataset: pink: PCNN, red: atlas-based, yellow: RATS, green: ground truth. All three methods produce decent results; however, the brainstem has a strong intensity artifact for this subject, causing problems for all methods to varying degrees. The “worst” subject was chosen as the subject that had the lowest Dice score average over the three methods.