Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session A

Poster G2 - Automatic Segmentation of Small Animal Brain Tumor MRI Using Deep Learning

Thursday, October 24, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

MG

Madison Galdieri

Undergraduate Research Associate
Wake Forest School of Medicine
Cornelius, North Carolina, United States

Co-Author(s)

Chad Arledge

Graduate Research Associate
Wake Forest University School of Medicine
Winston Salem, North Carolina, United States

Primary Investigator(s)

DZ

Dawen Zhao

Professor, Biomedical Engineering and Cancer Biology
Wake Forest University School of Medicine, United States

Introduction: Brain metastasis (BM) is the most common malignant brain tumor, accounting for approximately 40 percent of all brain cancers. The prognosis of BM is extremely poor, with a median survival of 4-6 months, even with aggressive treatment including radiotherapy and surgical resection [1]. BM is characterized by tumor cells infiltrating into the surrounding normal brain and the development of multi-focal tumor lesions. Thus, accurate demarcation of BM boundaries on radiographic MR images is imperative for treatment planning and consequently prognosis. Current traditional segmentation methods, including manual segmentation and thresholding, are largely time consuming and require extensive expertise. Recently, deep learning has emerged as a novel approach in image segmentation tasks, offering improved segmentation accuracy and efficiency compared to traditional methods. Indeed, several studies have revealed the utility of deep learning-based segmentation in clinical applications. However, these approaches have not been well studied and established in small animal brain tumor research. Segmentation in small animal BM MRI can be more challenging than human applications due to factors including animals’ smaller size, multi-focal lesion development, higher anatomical variability, and lower signal-to-noise ratio. In the present study, we have investigated the utility of deep learning for accurate and efficient segmentation of small animal BM MRI.

Materials and

Methods: Triple negative breast cancer brain metastasis (BM) MDA-MB-231-Br cells were injected into the left ventricle of nude mice (n = 25) under ultrasound guidance [2]. High-resolution T2-weighted images (T2-WI) and T1-contrast enhanced (T1-CE) images were obtained using a 7T Bruker BioSpin small animal MRI scanner following tumor formation (matrix size: 256 × 256; spatial resolution: 86 × 86 μm2). Tumor-bearing MRI images (n = 347 images; n = 1065 lesions) were manually segmented by an imaging expert to establish the ground truth. A U-Net neural network was trained to learn the relationship between small animal brain tumor MR images and the ground truth segments (Figure 1) [3]. Training was performed using k-fold cross validation allowing all animals to serve as the testing case in separately trained networks. The following hyperparameters were determined for achieving the best network performance: a training to validation ratio of 90:10, a mini-batch size of 2, an initial learning rate of 1e-4, and a maximum epoch of 300. A generalized dice loss function was employed during training of the network. Dice similarity coefficient, Jaccard index, sensitivity, and precision were quantified for each k-fold to assess the accuracy of the deep learning approach for small animal brain tumor segmentation. The accuracy of the U-Net deep learning approach in segmentation of T2-WI (n = 257) and T1-CE images (n = 90) was compared using a student's two-tailed t-test for the dice, Jaccard index, sensitivity, and precision.

Results, Conclusions, and Discussions: Approximately three weeks after tumor cell injection, hyperintense multi-focal tumor lesions were observed on T2-WI, as shown in Figure 2. T1-CE images revealed differential permeability between individual lesions, in which many lesions retained an intact blood-tumor barrier (BTB) as evidenced by no enhancement on T1-CE images. On average, manual segmentation of a single animal required times up to 15 minutes. In contrast, the developed U-Net deep learning approach was found to segment a single animal in less than a second. The U-Net deep learning approach accurately segmented individual BM lesions, with an average dice of 0.67 and a Jaccard index of 0.55 across all animals (Figure 3). It was also found that the deep learning approach had robust sensitivity and precision with values of 0.83 and 0.64, respectively. Importantly, the developed deep learning approach could accurately segment both T2-WI and T1-CE images, albeit different imaging sequences and features. As shown in Figure 3, the U-Net deep learning approach had significantly higher dice (p < 0.05) and Jaccard index (p < 0.01) for T1-CE images relative to T2-WI indicating the developed approach more accurately segments T1-CE images. Furthermore, it was found that the U-Net had similar sensitivity in both T1-CE and T2-WI, but significantly higher precision in T1-CE images (p < 0.01).

Although the U-Net approach segments small animal BM images with high accuracy, the deep learning approach had difficulty segmenting smaller metastases as well as lesions located near cerebrospinal fluid (CSF) as shown in Figure 2. We believe this to be a driving force for lower precision in T2-WI segmentations due to the presence of CSF in T2-WI leading to higher rates of false positives. Despite these challenges, we anticipate that through including additional testing cases, the developed approach will serve as a robust surrogate segmentation method in small animal brain tumor MRI. Future studies may include the incorporation of additional brain tumor phenotypes including glioblastoma.

Acknowledgements (Optional): This project was supported in part by the NIH/NLM (R25 LM014214) in the Department of Biomedical Engineering and Center for Artificial Intelligence Research at Wake Forest University School of Medicine and grant award R01 CA194578 from the NIH/NCI.

References:
[1] Moravan MJ et al. Cancer. 2020. 126(7): 1390-1406.
[2] Arledge CA et al. Cancers. 2023. 15(10), 2703.
[3] Ronneberger O et al. 2015. arXiv:1505.04597.