3D Segmentation of Unruptured Intracranial Aneurysms using Task Specific Transfer Learning and Pure Convnets

In this paper, we present a novel approach for segmentation of Unruptured Intracranial Aneurysms (UIAs) in Time-of-Flight magnetic resonance angiographies (TOF-MRAs). We propose, 1) a U-Net based pure convolutional neural network architecture (U-ConvNeXt) mimicing self attention and capturing global spatial information as in Vision Transformers, 2) a novel combined loss function of Dice Loss and modified Hausdorff Distance (Focal Hausdorff Loss) which optimizes region localization and penalizes for spatial similarity between ground-truth and prediction masks, 3) a task-specific Transfer Learning approach leveraging pre-trained models of Brain Tumour Segmentation on large datasets. We first train and evaluate our model on the ADAM dataset released by MICCAI which comprises of 113 TOF-MRA scans of 78 unique patients with 93 scans having at least one UIA. The aneurysms differ significantly in size, location and appearance with huge imbalance in background to foreground ratio making segmentation a highly challenging task. On the holdout set, we get a Dice Score Coefficient of 0.87 and a Hausdorff Distance of 1.91 from our best model which is 12% higher than the current SOTA. We also evaluate our model on another independent holdout set comprising of 38 TOF-MRA scans from Lausanne University Hospital and get a baseline Dice Score coefficient of 0.84 and a Hausdorff Distance of 2.15 for comparison with future works. The aneurysms with diameter size less than 3mm are the most challenging to segment with a Dice Score Coefficient of 0.76 and the simplest with size greater than 7mm with a Dice Score Coefficient of 0.95.