Deep Learning Enables Automatic Detection and Segmentation of Brain Metastases on Multi-Sequence MRI (1903.07988v1)

Abstract: Detecting and segmenting brain metastases is a tedious and time-consuming task for many radiologists, particularly with the growing use of multi-sequence 3D imaging. This study demonstrates automated detection and segmentation of brain metastases on multi-sequence MRI using a deep learning approach based on a fully convolution neural network (CNN). In this retrospective study, a total of 156 patients with brain metastases from several primary cancers were included. Pre-therapy MR images (1.5T and 3T) included pre- and post-gadolinium T1-weighted 3D fast spin echo, post-gadolinium T1-weighted 3D axial IR-prepped FSPGR, and 3D fluid attenuated inversion recovery. The ground truth was established by manual delineation by two experienced neuroradiologists. CNN training/development was performed using 100 and 5 patients, respectively, with a 2.5D network based on a GoogLeNet architecture. The results were evaluated in 51 patients, equally separated into those with few (1-3), multiple (4-10), and many (>10) lesions. Network performance was evaluated using precision, recall, Dice/F1 score, and ROC-curve statistics. For an optimal probability threshold, detection and segmentation performance was assessed on a per metastasis basis. The area under the ROC-curve (AUC), averaged across all patients, was 0.98. The AUC in the subgroups was 0.99, 0.97, and 0.97 for patients having 1-3, 4-10, and >10 metastases, respectively. Using an average optimal probability threshold determined by the development set, precision, recall, and Dice-score were 0.79, 0.53, and 0.79, respectively. At the same probability threshold, the network showed an average false positive rate of 8.3/patient (no lesion-size limit) and 3.4/patient (10 mm3 lesion size limit). In conclusion, a deep learning approach using multi-sequence MRI can aid in the detection and segmentation of brain metastases.

• The paper introduces a CNN-based method using a modified GoogLeNet architecture with 2.5D deep learning to accurately detect brain metastases on multi-sequence MRI.
• The research achieves high diagnostic performance with an average AUC of 0.98 and a Dice score of 0.79, while minimizing false positives with size threshold adjustments.
• The study underscores the clinical potential of AI in enhancing stereotactic radiosurgery planning and reducing manual workload through real-time, automated segmentation.

Automated Detection and Segmentation of Brain Metastases using Deep Learning and Multi-Sequence MRI

The paper "Deep Learning Enables Automatic Detection and Segmentation of Brain Metastases on Multi-Sequence MRI" presents a methodological advancement in computational radiology for the detection and segmentation of brain metastases using deep learning techniques. The paper employs a fully convolutional neural network (CNN) based on a modified GoogLeNet architecture, applying a unique approach with multi-sequence MRI data to achieve notably high accuracy in detecting brain metastases.

Methodology and Experimentation

The research encompasses a retrospective paper with data collected from 156 patients who had brain metastases originating from various primary cancers. Imaging protocols included pre-therapy MR images from different contrasts and a combination of pre- and post-gadolinium T1-weighted 3D fast spin echo (CUBE), post-gadolinium T1-weighted 3D axial IR-prepped FSPGR (BRAVO), and 3D CUBE fluid-attenuated inversion recovery (FLAIR). The ground truth for the metastases was manually annotated by neuroradiologists.

The CNN architecture employed in this paper, adapted to optimize for meta-data segmentation, integrated 2.5D deep learning to balance computational efficiency and segmentation accuracy. The model was trained and validated using a subset consisting of 100 training and 5 development patient cases, with performance evaluated over 51 test patients categorized based on the number of metastases (1-3, 4-10, >10).

Results

Empirical results showcase a high area under the ROC curve (AUC), averaging 0.98 across all test subjects, illustrating the robustness of the model to faithfully predict metastasis contours. The performance was marginally nuanced across different lesion categories, with AUC values of 0.99, 0.97, and 0.97 for patients with 1-3, 4-10, and more than 10 metastases, respectively. The Dice/F1 score for segmentation accuracy was 0.79, with false positive rates reported at 8.3 per patient without a lesion-size limit and reduced to 3.4 with a size limit of 10 mm³. Such data indicate optimal detection threshold calibration and mark significant clinical potential concerning patient diagnosis and treatment planning.

Discussion and Implications

This research contributes methodological insights into the utility of deep learning for radiological applications, specifically in the context of the multi-faceted challenge represented by brain metastases. By demonstrating successful segmentation of metastases, the paper opens up practical pathways to enhancing stereotactic radiosurgery planning, potentially diminishing manual workload and minimizing observer variability in clinical settings. Furthermore, this CNN model is agile enough for deployment on mobile GPUs, signaling a positive direction for patient-centric healthcare innovations, especially concerning real-time diagnostic capabilities.

Future Directions

Despite the promising outcomes, the paper acknowledges limitations, notably, the constrained single-center dataset and the necessity for all MRI contrast sequences for model effectiveness. The potential for model generalization across multi-site data and the exploration of alternative CNN constructions remain open fields for further research. Developing models that can accommodate variable imaging inputs could increase applicability across diverse clinical settings while enhancing overall robustness. Such advancements would align with the strategic goals of leveraging AI in precision medicine to improve patient-specific therapeutic strategies efficiently.

In summary, the presented research illustrates substantial progress in automating radiological procedures through deep learning, thereby contributing to the refinement of diagnostic workflows in healthcare systems.