Enhancing mitotic figure detection using attention
modules in digital pathology
May Hlaing Kyi
School of Computing and Engineering
Supervisors:
Dr Neda Azarmehr
School of Computing and Engineering
Professor Massoud Zolgharni
School of Computing and Engineering
Counting mitotic figures is essential for cancer grading and prognosis; however, manual counting is tedious and prone to pathologist discordance due to diverse appearances of mitotic figures. Developing an effective detection model is difficult due to the complex growth patterns and similarities to non-mitotic cells. Convolutional neural networks have been used to automate this process, yet current results are still inadequate for clinical use.
Aim and objective:
This study aims to investigate the potential of incorporating various state-of-the-art attention mechanisms into a ResNet backbone for differentiating mitotic figures from mimics in Whole Slide Images (WSIs). The objectives are to develop and evaluate the developed framework for the automated mitotic figure detection and to visualise detected mitotic figures at the WSI level.
Research questions:
How effectively do attention modules enhance the accuracy of mitotic figure detection in WSI?
Research design and method:
This study employs a detection and classification task (mitosis versus mimics) based on the RetinaNet model. We designed and trained the RetinaNet models with several state-of-the-art attention modules including the Convolutional Block Attention Module (CBAM), Squeeze-and-Excitation (SE), Enhanced Channel Attention, and Hybrid Attention Module integrated into ResNet50 backbone. All developed models were evaluated using the public Canine Mammary Carcinoma dataset. The performance of models is evaluated using several metrics, including F1-score, precision, and recall, across test-set.
Findings:
Results indicate that the models with a ResNet50 backbone incorporating CBAM and SE attention mechanisms achieved F1 scores of 0.793 and 0.784, respectively, outperforming the standard RetinaNet architecture.
Implications of the findings:
Utilising attention mechanism can improve the accuracy of mitosis detection model. This improvement leads to more accurate diagnoses, reduced variability, and ultimately personalised treatment plans.