RESEARCH Open Access© The Author(s) 2025.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the.
You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by.
Colorectal cancer (CRC) is currently the third most com - mon cancer in men and the second in women, as well as the second leading cause of all deaths from cancer [ 1]. Incidence and mortality rates differ across regions, which can be partially attributed to economic inequalities and differences in healthcare access [ 2, 3].
While effective screening measures has fallen the incidence of CRC in France, it has markedly increased in the French Carib - bean islands during the past twenty years and continues to increase in women [ 4].
Colonoscopy is effective in preventing CRC and low - ering its mortality rate (up to 60%), by detecting all BMC Medical Informatics and Decision Making *Correspondence: Moana Gelu-Simeon moana.simeon@chu-guadeloupe.fr 1Service d’Hépato-Gastroentérologie, CHU de la Guadeloupe, Pointe- à-Pitre F-97100, France 2Univ Antilles, Univ. Rennes, INSERM, EHESP , IRSET (Institut de Recherche en Santé,.
Selection and training process of the model First step: Architecture application design We identified a global architecture application enable to real-time instance segmentation with three frameworks method architectures: YOLACT and YOLACT + + based on “one-stage architecture” , and Mask-RCNN on “two- stage architecture” .
The YOLACT and YOLACT + + mod - els were initialized with weights pretrained on the ImageNet dataset [ 36], consistent with the approach reported in the original YOLACT + + implementation by Bolya et al. [ 27], Mask-RCNN was initialized with COCO database [ 37].
These method architectures with instance function - ality were associated with three different backbones: ResNet-50, ResNet-101 and DarkNet-53 to extract fea - tures of the data images. These backbones are most commonly adopted in the literature in association with YOLACT or YOLO [ 15, 26, 27]. Beyond algorith - mic modifications, diverse architectural structures like ResNet or VGG for YOLACT and.
The validated model was achieved with the best mAP score for bounding IoUs ranging from 0.5 to 0.95 and FPS > 30 (Fig. 2). It consisted of the YOLACT method plus the ResNet-50 backbone, with FPS = 32.82, image size = 550, mAP = 72.32 and APs associated with IoUs of 50, 75, and 95% (Table 2).
The validated model (YOLACT, ResNet-50), named “RTPoDeMo” , on a parallel architec - ture computer (NVIDIA Tesla P100 GPU and Intel Xeon 2.20 GHz CPU) was then connected to the endoscopy processor to enable the visualization of polyp delineation during the colonoscopy procedure (Fig. 4).
Nineteen colonoscopies were performed between 1st June and 9 July 2021, one colonoscopy had been excluded because carried out in the context of lower gastroin - testinal bleeding and 18 unaltered colonoscopy videos were analyzed. During the 18 colonoscopy procedures, the senior gastroenterologists identified 29 polyps. The median video duration was 37 min, with a minimum duration of 7 min and a.
This study evaluated the performance of YOLACT derived RTPoDeMo designed for real-time application on prospectively recorded colonoscopy videos. We have reported here all the steps that led to the selection and validation of the model, including its design, the prepa - ration of medical images, the training process and the benchmarking of each model and evaluation of their performances.
We have demonstrated that RTPoDeMo deep-learning model was able to perform, real-time polyp detection and instance segmentation, but also the delineation of colorectal polyps by means of green fluo - rescence. This preliminary study, conducted on 18 pro - spectively recorded colonoscopy videos, achieved high levels of per-image specificity, sensitivity and accuracy, and good agreement with the.
Concerning the delineation of polyps, various seg - mentation techniques are available. These encompass approaches such as considering the entire scene (seman - tic segmentation), identifying individual objects (instance segmentation), amalgamating both strategies (panoptic segmentation), relying on object borders (boundary- based segmentation), or segmenting the image into sections (hierarchical.
We found high performance of RTPoDeMo with a novel approach of multi-instance segmentation, real-time polyp detection and delineation. Its feasibility in clini - cal practice is supported by its short processing time, so that it could be easily employed during colonoscopy procedures with negligible latency and without slowing exploration of the colon.
This model displayed clinical applicability and simplicity of use. The innovative value of our system relies on a precise delineation of a polyp, which could provide more precision about the character - istics of the polyp. This will be more accurate to evaluate the colorectal polyps before resection with a lower risk of missing a part of the lesion.
These potential values in recognizing the shape and the pathology of colorectal lesions should be improved in further studies. We hope to confirm the good results obtained with this model in a future experiment on our regional image databases collected as part of a health data warehouse, in order to evaluate the model’s performance in specific populations. Abbreviations RTPoDeMo Real-time Polyp.