Title
Artificial Intelligence–Assisted Diagnosis of Pulmonary Pathological Changes in Magnetic Resonance Imaging
Authors
[ 1 ] Wydział Techniczny, Akademia im. Jakuba z Paradyża | [ P ] employee
Scientific discipline (Law 2.0)
Year of publication
2025
Published in
Article type
scientific article / paper
Publication language
english
Keywords
EN
- Artificial Intelligence (AI)
- Deep Learning (DL)
- Pulmonary MRI
- Medical Image Segmentation
- Hybrid U-Net Architecture
Abstract
EN This article presents a scientific contribution from Phase II of the “COVID FAST” Project, which aimed to develop an artificial intelligence (AI) system to support the diagnosis of post-COVID-19 pulmonary complications based on magnetic resonance imaging (MRI). The study reports the development of a hybrid deep neural network (DNN) model for the automatic segmentation of pathological changes in the lungs. The model integrates U-Net architectures with both multi-label and logits-based outputs, enhancing detection performance – particularly for rare and morphologically diverse pathological lesions. The model achieved high accuracy in identifying pulmonary abnormalities, with mean IoU values exceeding 90% for four relatively small and infrequent lesion types (e.g., adhesions/fibrosis); between 91% and 93.6% of test cases in each of these categories reached IoU > 70%. For the more common lesion type – ground-glass opacity – the target performance level was not achieved (62.8% of test cases with IoU > 70%; mean IoU = 77%). A detailed annotation analysis revealed low inter-observer agreement for ground-glass opacity (59.1% of cases with IoU > 70%), whereas the AI model achieved slightly higher consistency (62.8%). Furthermore, the analysis of false positive test cases (15%) for ground-glass opacity indicated that the model frequently identified lesions that had been overlooked by experts in the reference images. The study results highlight the potential of AI-assisted analysis in lung MRI interpretation and provide a strong rationale for continued research within the Project, as well as for the model’s future clinical applications in healthcare.
Pages (from - to)
5270 - 5279
License type
CC BY-NC-ND (attribution - noncommercial - no derivatives)
Open Access Mode
open repository
Open Access Text Version
final published version
Release date
11.2025
Date of Open Access to the publication
at the time of publication
Ministry points / journal
5
Ministry points / conference (CORE)
70
This website works thanks to the software SINUS
created by Poznan Supercomputing and Networking Center
Log in to submit to SIS: