This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.
The power of accurate segmentations in AI-driven solutions for medical imaging
The segmentation process involves precisely identifying and delineating a region of interest (ROI) within an image. In medical imaging, these ROIs may include lesions, organs, healthy tissues, or structures, delineated in one or multiple modalities, such Magnetic Resonance Imaging (MRI), Computerized Tomography (CT) or Positron Emission Tomography (PET), among others.
Precision at Quibim
At Quibim, the precision of our segmentation results is pivotal. It serves as a sophisticated tool not only for characterizing the aggressiveness and heterogeneity of tumors linked to patient evolution but also as the cornerstone for developing advanced AI-based algorithms, capable of making predictions at both lesion and patient levels. For this reason, a standardized segmentation process is systematically implemented in every study we manage at our company, leading to the definition of reproducible and standard protocols refined to each specific use case. Not only the segmentation of a lesion is important, but also its surrounding organs and tissues. As an example, for the overall survival (OS) prediction in lung cancer, not only the lesion ROI might be used, but also other structures nearby such as lungs, heart, bone/calcium, liver, metastasis, among others.
Challenges in segmentation
There are many factors that significantly affect the segmentation and image reading, like spatial and contrast resolution, image noise as well as variability in the shape, texture and perception of pathologies1. This highlights the importance of conducting quality controls before embarking on any image processing, such as the segmentation that needs to be introduced to the training of AI models as a label or annotation. It consists of a set of tasks to select images that have enough diagnostic quality and are suitable to be segmented. Some of the tasks include the selection of the best sequence to perform the segmentation according to criteria such as spatial and contrast resolution, signal-to-noise ratio, absence of artifacts, and coverage of ROI, among others. After this quality control, the next phase is the segmentation and obtention of the ROI. The beauty of this segmentation lies in its voxel-by-voxel execution, providing three-dimensional insights into the ROI. This approach surpasses conventional criteria that rely on diameter measurements, thereby capturing a more comprehensive and nuanced representation of the tissue under study.
Quality control and segmentation process
Quibim decided to address this crucial step at AI algorithms manufacturing by having a dedicated team responsible for meticulous quality checks and precise segmentations. This ensures strict adherence to the highest quality standards directly influencing the performance of our AI-empowered automated segmentation tools, as well as predictive and prognostic algorithms. Serving as the inception point, these processes mark the initial interaction with the imaging exams, setting the stage for achieving success in the development of Quibim products.
In addition to the above purpose, these accurate and validated segmentation serves as a highly valuable source of databases. This wealth of data empowers our company to further advance, developing proprietary automatic segmentation algorithms guided by cutting-edge AI technology. This strategic utilization enhances our capacity for innovation and ensures a robust foundation for the continuous evolution of our QP-solutions.
1-Segmentation of the central zone, peripheral zone of the prostate and seminal vesicles on Axial T2-wheited magnetic resonance imaging.
2-Whole body 18-FDG PET/TC imaging exam of a patient with DLBCL in the coronal plane(left) and segmentation mask where every lesion is a unique label (right).
3, 4-Segmentation of the left lung tumor mass on axial Computerized Tomography slices with 3D reconstruction.
5-Segmentation of various anatomical structures in a coronal plane on a Computerized Tomography study.
References
- deSouza NM, van der Lugt A, Deroose CM, et al. Standardised lesion segmentation for imaging biomarker quantitation: a consensus recommendation from ESR and EORTC. Insights Imaging. 2022;13(1):159. Published 2022 Oct 4. doi:10.1186/s13244-022-01287-4
Authors
Raúl Yebana – Image Analysis Manager
Víctor Climent – Image Analysis Technician