Detection of Brain tumors using integrated Mask-RCNN and spatial attention mechanism.

DOI: https://doie.org/10.1111/Jbse.2024427228

Chandrakant M. Umarani, S. G. Gollagi, Shridhar Allagi, Kuldeep Sambrekar, Sanjay Ankali

Keywords:

Brain tumor type, Mask RCNN, Spatial attention mechanism, YOLOv8..

Abstract:

The proposed novel deep learning methodology incorporating Mask-RCNN integrated with spatial attention mechanism precisely identifies three basic brain tumors types such as glioma, meningioma, and pituitary tumors by concentrating on pertinent spatial areas within an MRI scan, thereby enhancing the detection of tumor types from MRI images. The proposed leightweight approach attained a mAP of 98.96%, surpassing the baseline Mask RCNN model which achieved a mAP of 88.34%, in contrast to YOLOv8's 89.30% mAP for a custom dataset comprising 1322 annotated MRI scans. These scans were annotated utilizing the VGG annotator tool, and the model was trained for 40 epochs with a batch size of 35, making use of the computational capabilities of a T4 GPU. The integration of a spatial attention mechanism into Mask-RCNN demonstrates superior performance in comparison to both the Basic Mask-RCNN and YOLOv8 in terms of average loss, accuracy, and mean average precision (mAP) for the identification of glioma, meningioma, and pituitary brain tumor subtypes which can be the lightweight model for clinical integration to accurately detect early tumors. 

PDF