Tolga D. Dittrich, MD
Nielsen M, Waldmann M, Frölich AM, Flottmann F, Hristova E, Bendszus M, Seker F, Fiehler J, Sentker T, Werner R. Deep Learning-Based Automated Thrombolysis in Cerebral Infarction Scoring: A Timely Proof-of-Principle Study. Stroke. 2021.
The success of mechanical thrombectomy is commonly measured by the TICI (Thrombolysis in Cerebral Infarction) score. The score is determined by the visual assessment of the digital subtraction angiography (DSA) images during the intervention by the treating interventionalist. Despite modifications of the original scale (e.g., modified TICI [mTICI]) that have become established in the meantime, a relatively high inter- and intraobserver variability of TICI scores can be observed. This investigator dependency poses a challenge, particularly regarding clinical studies, as it may affect the comparability of results.
In the present study, based on occlusions of the middle cerebral artery in the M1 segment, Nielsen and colleagues sought to develop an automated and thus more objective TICI assessment using a deep learning (DL) approach. Agreement between DL and expert-based assessment (gold standard) was evaluated, and a comparison with corresponding published numbers on expert assessment variability was performed.
For the two DSA datasets used from different centers, DL-based predictions and expert-assigned mTICI scores showed agreement in the range of published intra- and inter-rater variability. The dichotomous distinction between open (mTICI score ≥2b) and occluded vessel (mTICI score <2b) was achieved reliably. Focusing on the individual mTICI subclasses, the prediction performance was lowest for mTICI 2b (for all experiments).
The figure illustrates the accuracy assessment between the network-based prediction and the expert assessment of the mTICI scoring (results after updated model training with external data [i.e., experiment III], adapted from the figure in the original publication).
DL-based automated TICI scoring emerges as a promising approach for the future. DL methods seem particularly well-suited for managing large and complex imaging data as they can learn automatically from given training data sets. Another significant advantage of an automated evaluation is that it is investigator-independent, increasing comparability in clinical trials. The application of such DL techniques to larger data sets and other occlusion sites is eagerly awaited.