Sub-Object Data Augmentation And Custom Loss Functions for Automated Anatomical Landmarks Localisation in 3D Brain MR Images

In medical image analysis, the landmarks detection helps to localize each anatomy for tasks like image registration. Deep learning models help to automate the landmark detection pipeline reducing medical technician efforts. The major challenge faced to apply deep learning in medical image analysis is the limited data available for training these models. In this work, we implemented several approaches for data augmentation to address this problem. We modified 2D Flat-net to 3D Flat-net landmark detection network, later used sub-object data augmentation and custom loss function - Masked loss and Soft-argmax loss. We trained and evaluated the combination of data augmentation and custom loss function. On the test set (n=63), using 3D Flat-net with sub-object augmentation, and masked loss function performed the best with Mean Absolute Error improved by 43 % (from 9.84 mm to 5.76 mm), Mean Euclidean Error improved by 34 % (19.94 mm to 13.16 mm). Root Mean Square Error (RMSE) improved by 39 %(from 11.52 mm to 7 mm) compared to Soft-argmax loss function and no augmentation. Data Augmentation of sub-objects and masked loss function improved 3D Flat-net performance for localizing landmarks on medical images.