International audience

While the fine-tuning process of extensive contextual language models often demands substantial computational capacity, utilizing generic pre-trained models in highly specialized domains can yield suboptimal results. This paper aims to explore an innovative approach to derive pertinent word embeddings tailored to a specific domain with limited computational resources (The introduced methodologies are tested within the domain of hepatic surgery, utilizing the French language.). This exploration takes place within a context where computational limitations prohibit the fine-tuning of large language models. A new embedding (referred to as FTW2V) that combines Word2Vec and FastText is introduced. This approach addresses the challenge of incorporating terms absent from Word2Vec’s vocabulary. Furthermore, a novel method is used to evaluate the significance of word embeddings within a specialized corpus. This evaluation involves comparing classification scores distributions of classifiers (Gradient Boosting) trained on word embeddings derived from benchmarked Natural Language Processing (NLP) models. As per this assessment technique, the FTW2V model, trained from scratch with limited computational resources, outperforms generic contextual models in terms of word embeddings quality. Additionally, a computationally efficient contextual model rooted in FTW2V is introduced. This modified model substitutes Gradient Boosting with a transformer and integrates Part Of Speech labels.

The segmentation of the liver, lesions, and vessels from pre-operative CT scans is of major importance in hepatic surgery planning. However, large databases with reference segmentations for these regions of interest remain unavailable, a challenge often encountered in medical image segmentation. In this work, we propose the FuSe loss, a novel loss function for multi-task learning on datasets with partial annotations. By employing the nnU-Net’s 3D self-configuring pipeline to calibrate and train a deep network for the joint segmentation of the liver, lesions, and vessels, we show how the FuSe loss allows to learn from the differently annotated IRCAD and LiTS datasets, improving the overall baseline segmentation performance. With the FuSe loss, the dice scores reached up to 95.9%, 70.6% and 60.0% for the liver, lesions, and vessels respectively.

This paper introduces an Augmented Reality (AR) system for open liver surgery. Although open surgery remains the gold-standard for the treatment of complex tumors and central lesions, technological issues actually prevent using AR with sufficient accuracy for clinical use. We propose a markers-based method allowing for the tracking and the deformation of a preoperative model in real-time during the surgery. Markers are manually placed on the surface of the organ after opening the abdominal cavity, and tracked in real-time by a set of infrared cameras. Our framework is composed of both a nonrigid initial registration method, providing an estimation of the location of the markers in the preoperative model, and a realtime tracking algorithm to deform the model during the surgery (even for large deformation or partial occlusion of the organ). The method is validated on both synthetic and ex-vivo samples; in addition, we demonstrate its applicability in the operating room during a liver resection surgery on a human patient. Preliminary studies provided promising results to improve the location of tumors, and to help surgeons into planning the ideal resection intraoperatively.