Session Item

Deep Learning (DL) models for tumor autodelineation provide significant advantages but expert validation and correction of model predictions remains critical for both scientific and clinical use. The question of optimal AI-expert interaction therefore should be crucial for radiooncology.

An interactive pipeline was created to accelerate segmentation of MRI brain metastases (BM) datasets for scientific use and to serve as blueprint for future AI-assisted clinical workflows. The pipeline incorporates 2 types of AI-expert interaction: 1) sequential use of DL models for subtasks with intermediate expert interaction 2) an interactive DL model with dedicated input channels for user interaction. The pipeline contains 5 steps: 1) preprocessing and autosegmentation of T1ce-MRI datasets with a high-specificity and high-sensitivity DeepMedic+ (DM+) model for identification of high-confidence BMs and BM candidates, respectively; 2) expert review of BM candidates; 3) expert identification of false negative BMs; 4) nnU-net  autosegmentation of identified BMs and combination with prior DM+ autosegmentations and 5) interactive correction of BM autosegmentations with an interactive DL model and expert scribbles. The DM+ model had been trained on 600 expert delineated whole-brain BM datasets, while the nnU-net model (3D fullres) was trained on 1875 subvolumes of 6x6x6 cm³ centered on each BM derived from the same dataset. The interactive DL model (nnU-net, 3D fullres) was trained to correct a given BM segmentation (predicting 2 classes: add+remove) using the 1875 MRI subvolumes and 37500 simulated triplets of corrupted BM segmentations and user scribbles as input for training. The interactive pipeline is implemented as Python module for 3DSlicer and was evaluated on a set of 20 independent datasets containing 61 manually delineated BMs.

Mean expert time required for each BM dataset using the interactive workflow was 04:35 min (range, 00:58–10:12). 60 of all 61 BMs in the test set were identified using the interactive workflow (sensitivity, 98.4%) with the single missing lesion having been predicted by DM+, but subsequently disregarded by the interacting expert. Interestingly, 9 additional BMs, missed by manual delineation, were detected with the interactive AI-assisted workflow in the test set. Regarding segmentation accuracy, mean surface Dice score (sDSC) for BMs was 0.94 (0.62–1.00) with significant improvement because of the interactive correction step (mean +0.012, p=0.001).

We present an interactive open-source pipeline for BM autodelineation in MRI datasets with multiple AI-expert interactions that accelerates the creation of annotated BM MRI datasets for scientific use and exemplifies AI-expert interaction for potential future clinical workflows.