Beyond rankings: Learning (more) from algorithm validation.

Challenges have become the state-of-the-art approach to benchmark image analysis algorithms in a comparative manner. While the validation on identical data sets was a great step forward, results analysis is often restricted to pure ranking tables, leaving relevant questions unanswered. Specifically, little effort has been put into the systematic investigation on what characterizes images in which state-of-the-art algorithms fail. To address this gap in the literature, we (1) present a statistical framework for learning from challenges and (2) instantiate it for the specific task of instrument instance segmentation in laparoscopic videos. Our framework relies on the semantic meta data annotation of images, which serves as foundation for a General Linear Mixed Models (GLMM) analysis. Based on 51,542 meta data annotations performed on 2,728 images, we applied our approach to the results of the Robust Medical Instrument Segmentation Challenge (ROBUST-MIS) challenge 2019 and revealed underexposure, motion and occlusion of instruments as well as the presence of smoke or other objects in the background as major sources of algorithm failure. Our subsequent method development, tailored to the specific remaining issues, yielded a deep learning model with state-of-the-art overall performance and specific strengths in the processing of images in which previous methods tended to fail. Due to the objectivity and generic applicability of our approach, it could become a valuable tool for validation in the field of medical image analysis and beyond.

Deep learning-based classification of DSA image sequences of patients with acute ischemic stroke.

PURPOSE: Recently, a large number of patients with acute ischemic stroke benefited from the use of thrombectomy, a minimally invasive intervention technique for mechanically removing thrombi from the cerebrovasculature. During thrombectomy, 2D digital subtraction angiography (DSA) image sequences are acquired simultaneously from the posterior-anterior and the lateral view to control whether thrombus removal was successful, and to possibly detect newly occluded areas caused by thrombus fragments split from the main thrombus. However, such new occlusions, which would be treatable by thrombectomy, may be overlooked during the intervention. To prevent this, we developed a deep learning-based approach to automatic classification of DSA sequences into thrombus-free and non-thrombus-free sequences. METHODS: We performed a retrospective study based on the single-center DSA data of thrombectomy patients. For classifying the DSA sequences, we applied Long Short-Term Memory or Gated Recurrent Unit networks and combined them with different Convolutional Neural Networks used as feature extractor. These network variants were trained on the DSA data by using five-fold cross-validation. The classification performance was determined on a test data set with respect to the Matthews correlation coefficient (MCC) and the area under the curve (AUC). Finally, we evaluated our models on patient cases, in which overlooking thrombi during thrombectomy had happened. RESULTS: Depending on the specific model configuration used, we obtained a performance of up to 0.77[Formula: see text]0.94 for the MCC[Formula: see text]AUC, respectively. Additionally, overlooking thrombi could have been prevented in the reported patient cases, as our models would have classified the corresponding DSA sequences correctly. CONCLUSION: Our deep learning-based approach to thrombus identification in DSA sequences yielded high accuracy on our single-center test data set. External validation is now required to investigate the generalizability of our method. As demonstrated, using this new approach may help reduce the incident risk of overlooking thrombi during thrombectomy in the future.