Automatic pancreas anatomical part detection in endoscopic ultrasound videos

ABSTRACTNowadays ranked 3rd, pancreatic cancer is predicted to become second highest mortality cancer in next decade. Early diagnosis is the only way to significantly improve survival rate for which endoscopic ultrasound (EUS) stands as the only viable medical imaging option. One challenging aspect is the complex interpretation of images during examination. It is not rare for non-experts to miss the screening of parts of the pancreas, leaving tumours undetected. Here, we propose an automated method to support non-expert clinicians in their practice by providing a deep-learning based tool able to detect the anatomical parts seen under EUS. We have collected 41 EUS videos and annotated the anatomy in each video frame. Considering the challenging and novel nature of EUS data, we propose a systematic analysis of feature extractors and temporal modules. We extend popular models with LSTM modules and compare their performance to the newly introduced vision transformers, yielding an overall comparison of 35 models. The results highlight the benefits of transformers for their ability to capture more anatomical context thanks to the division into patches and positional embeddings. As a result, our study paves the way to AI-assisted pancreas examination for early cancer detection.