doi: 10.18178/ijml.2024.14.4.1167
Multiscale Feature Learning and Cross Spatial Attention in Mask R-CNN for enhanced Cell Instance Segmentation
Dipankar Jiwani1
,
Rachita Saha1 ,
Isha Sehrawat1 ,
Anubha Gupta1,* ,
Anish Jain1 ,
Ritu Gupta2
- 1SBILab, Department of ECE, IIIT-Delhi, India
- 2Dr. BRA.IRCH, Laboratory Oncology Unit, AIIMS, New Delhi, India
- Manuscript receivedFebruary 29, 2024
- revisedJune 9, 2024
- acceptedJune 11, 2024
- publishedOctober 16, 2024
Abstract
Cell instance segmentation in medical imaging is pivotal for advancing diagnostics and treatment. Despite the acknowledged importance of Mask R-CNN for this task, we observed challenges in effectively distinguishing some boundary pixels, particularly in scenarios where cells are in close proximity. To address these issues, this research introduces three key enhancements: 1) Multiscale Feature Learning (MSFL), 2) the Cross Spatial Attention Module (CSAM), and 3) Novel training of UNet for guiding the training of Mask RCNN through CSAM module. MSFL utilizes various features produced by the FPN backbone across different scales, thereby minimizing data loss and enhancing the overall representation of the region of inter- est. The lightweight CSAM significantly enhances segmentation results by harnessing the inherent segmentation capabilities of U-Net in the medical domain. This novel approach not only rectifies boundary errors, but also enhances accuracy and robustness in medical image analysis. Importantly, the adaptable CSAM seamlessly integrates into various models, increasing segmentation accuracy without a substantial impact on the model size. The efficacy of this approach is demonstrated through its application on two distinct cell segmentation datasets. Results demonstrate a notable increase of 2.66% mean Intersection over Union (mIOU) from the baseline on the SegPC dataset and a significant improvement by 1.86% in the mAP@[0.5:0.95] on the Yeast Cell dataset.
Keywords
- medical imaging
- cell instance segmentation
- Mask R-CNN
- U-Net
- spatial attention
How to Cite
Dipankar Jiwani, Rachita Saha, Isha Sehrawat, Anubha Gupta, Anish Jain, and Ritu Gupta, "Multiscale Feature Learning and Cross Spatial Attention in Mask R-CNN for enhanced Cell Instance Segmentation," International Journal of Machine Learning, vol. 14, no. 4, pp. 113-118, 2024. https://doi.org/10.18178/ijml.2024.14.4.1167
Copyright & License
Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).