Laboratory of Medical Information Processing

Retinal diseases are among the leading causes of visual impairment worldwide, where timely diagnosis and management are critical to prevent irreversible vision loss and blindness, especially in regions with limited access to ophthalmologists. While artificial intelligence (AI) has shown remarkable potential for screening ocular diseases, many existing models are heavily dependent on the datasets used for their development and often experience significant performance drops when tested on external datasets. These limitations in robustness and generalizability reduce their practical applicability in clinical settings. In this study, we proposed our newly developed deep learning architecture, FlexiVarViT is designed to enhance robustness and generalizability by addressing the unique characteristics of optical coherence tomography (OCT) imaging, such as variable data (e.g., number of slices, resolution), while processing B-scans at their native resolution (i.e., without resizing) to preserve fine anatomical details. Our method was evaluated on three datasets for the accurate detection and classification of multiple retinal pathologies using OCT images. These datasets represent diverse imaging devices (Spectralis and Optovue) and populations from France, Russia, and Iran. The results highlight the versatility, robustness, and generalizability of our model across various patient demographics and imaging systems, demonstrating significant improvements compared to state-of-the-art methods and emphasizing its potential for clinical application.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-19286-y.

In low-dose computed tomography (LDCT) denoising tasks, it is often difficult to balance edge/detail preservation and noise/artifact reduction. To solve this problem, we propose a dual convolutional neural network (CNN) based on edge feature extraction (Ed-DuCNN) for LDCT. Ed-DuCNN consists of two branches. One branch is the edge feature extraction subnet (Edge_Net) that can fully extract the edge details in the image. The other branch is the feature fusion subnet (Fusion_Net) that introduces an attention mechanism to fuse edge features and noisy image features. Specifically, first, shallow edge-specific detail features are extracted by trainable Sobel convolutional blocks and then are integrated into Edge_Net together with the LDCT images to obtain deep edge detail features. Finally, the input image, shallow edge detail, and deep edge detail features are fused in Fusion_Net to generate the final denoised image. The experimental results show that the proposed Ed-DuCNN can achieve competitive performance in terms of quantitative metrics and visual perceptual quality compared with that of state-of-the-art methods.

International audience

Patient Specific Instruments (PSIs) have been introduced into the surgical workflow as a modern way to assist the surgeon in performing femur and tibia resection in Total Knee Arthroplasty (TKA). These PSIs are based on an accurate reconstruction of the surface of the knee’s bones.In this work, we propose two 3D-3D image-based registration methods to reconstruct an extended field-of-view of the knee joint using only a motorized ultrasound transducer. Those methods are: (1) a dense voxel-based registration method, which needs to preprocess the ultrasound images and form an ultrasound volume. Then, computing the Mutual Information (MI) for each relative displacement to align every pair of volumes, (2) a sparse point-based registration method, which takes into account the point set located on the surface of the bone in ultrasound images. This method detects bony features using ORB detector and matches the corresponding points to find the best transformation using Coherent Point Drift (CPD).The preliminary qualitative results performed in vitro show that from a set of consecutive ultrasound volumes, an extended field-of-view can be reconstructed using only ultrasound images without any external trackers. Results of the voxel-based approach show that MI is more robust against noise comparing to other similarity measures. On the other hand, results of point-based approach show that is much faster in computation with a low false-positive rate compared to other feature-detectors like SIFT and SURF. Furthermore, experiments show that CPD is less affected by noisy data compared to the classical ICP, which is promising to continue evaluating our work in vivo.

Background: To evaluate the occurrence of erectile dysfunction at 3 years (3yED) after prostate brachytherapy (BT) and to predict 3yED after treatment based on patients and treatments characteristics.Material and methods: From September 2007 to July 2015, 117 men with mild or no ED [International Index of Erectile Function (IIEF-5) > 16] underwent 125Iodine real-time ultrasound-guided low-dose rate BT to a total dose of 160 Gy for low-risk or favorable intermediate-risk prostate adenocarcinoma, and were followed prospectively during 3 years. Median age was 63 years (51–79). The post-implant dosimetric parameters on the postoperative computer tomography were derived from the dose–volume histogram of the prostate and the penile bulb (PB), crura, neurovascular bundles (NVBs) and internal pudendal arteries (IPAs). Potential clinical confounding factors were collected. Additionally, anatomical indexes reflecting the prostate anatomical location within the pelvis were studied. These variables were compared between patients with and without 3yED. 3yED was defined as an IIEF-5 score change to the lower category between baseline, with or without medication.Results: The 3yED rate was 59% (62% maintained an IIEF-5 > 16). On multivariate analysis, prostate D90% (p > .5) and pretreatment characteristics including age (p > .5), pre-implant potency (p > .5), diabetes (p = .08) and high cardiovascular risk rates (p = .1) did not influence the occurrence of 3yED. Only the PB dose especially the D10% > 51 Gy was associated with 3yED (p = .005). Conversely, dose to the crura, IPAs or NVBs did not seem to impact the erectile function. The prostate position, especially the apex location varied significantly between potent and impotent patients and 3yED was significantly associated with close position of the prostate apex to PB (p = .008).Conclusion: The most predictive factor of 3yED was the dose to the PB. This may be explained by variation in individual patients’ anatomy and this could allow for the development of better strategies to prevent ED.