Clinically acquired new challenging dataset for brain SOL segmentation: AJBDS-2023

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Neuro-oncology

Background:

• Space-occupying lesions (SOL) on brain MRI include benign and malignant tumors, necessitating accurate detection and segmentation.
• Existing brain tumor segmentation algorithms require clinically acquired datasets for real-time image analysis.
• Reducing inter-observer variability and radiologist workload in MRI reporting is crucial for efficient clinical practice.

Purpose of the Study:

• To introduce the AJBDS-2023, a novel dataset of multiplanar, multi-sequential MRI slices (MPMSI) for brain tumor detection.
• To facilitate the development of computer-aided detection (CAD) systems for unbiased and expedited MRI reporting.
• To provide a realistic platform for evaluating segmentation algorithms using preprocessed clinical data.

Main Methods:

• The AJBDS-2023 dataset comprises 10,667 clinically acquired, multiplanar, multi-sequential MRI slices (320x320x3) without preprocessing.
• Ground-truth annotations for tumor core and edema were manually created for 6,334 slices under radiologist supervision.
• A novel U-Net segmentation model was quantitatively assessed on 4,333 images from the AJBDS-2023 dataset.

Main Results:

• The U-Net algorithm trained on AJBDS-2023 achieved a precision of 77.4%, Dice Similarity Coefficient (DSC) of 82.3%, specificity of 87.4%, and sensitivity of 93.8%.
• The model demonstrated a confidence interval of 90.4%.
• The AJBDS-2023 dataset, with its un-preprocessed images, presents a more challenging and realistic evaluation environment.

Conclusions:

• The AJBDS-2023 dataset supports the development of advanced algorithms for brain tumor segmentation and detection.
• The proposed dataset enhances the realism in evaluating segmentation models, leading to more reliable computer-aided detection tools.
• Utilizing this dataset can significantly aid radiologists in improving the accuracy and efficiency of MRI brain tumor reporting.