Development of AI-Based Approaches for Automated Real-Time Detection of Endometriosis Lesions Using Endoscopic Image and Video Data

In the field of endometriosis, artificial intelligence (AI) has been used for diagnoses or even predictions of endometriosis before confirmation through laparoscopy. AI's significant potential in minimally invasive surgery lies in automatic image analysis, aiding in the detection of structures or anomalies based on image data. This offers the potential to detect endometriosis lesions during laparoscopy regardless of the indication. Training and creating such AI models are done using machine learning algorithms based on annotated data. These training data consist of image data with pixel-level annotations of the content that the model should detect. Deep learning (DL) algorithms have proven effective in image analysis, relying on neural networks to autonomously fill them with the most critical decision criteria for correct analysis of the image content. The trained AI model can then be applied to unknown data, providing the probability of detecting a structure for each pixel. Possible visual outputs of the model include outlining the detected content or segmenting, assigning predefined content to each pixel. The quality of the model depends crucially on a sufficiently large number and quality of training data. Quality includes correct annotation of data to prevent the model from learning errors. Diversifying image data by including negative examples in the training and test datasets is equally important. The F1-score is used as a measure of the model's quality, combining precision (P) with recall (R) to a value between 0 and 1, based on an annotated test dataset.

The goal is to achieve a high F1-score through the selection of training data and an appropriate DL algorithm. Parameters like image preparation optimization or DL algorithm parameters such as selecting different neural networks can improve the F1-score. The number of required training data for a good AI model depends on the complexity of the question and the number of contents to be detected, as the model can only recognize learned content. It is possible to iteratively adjust the selection of training data for different questions based on the achieved F1-scores after each training and testing. If necessary, the number of training data can be increased, and problematic image data, such as missing annotations, can be identified and corrected based on the results.