HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation
Master Thesis

Comparaison d'une analyse automatique d'IRM cérébrale par PIXYL versus imagerie de fusion dans le suivi de la sclérose en plaques

Abstract : The MRI evolution of white matter lesions is a good marker for the control of multiple sclerosis (MS) by treatments with a significant threshold of three. However, the interpretation of MRI requires time-consuming expertise and becomes more complex with the importance of the lesion burden. Semi-automatic methods such as image FUSION can help the radiologist detect active lesions. Automated segmentation methods such as PIXYL are being developed to characterize lesions objectively and quickly. The main objective of this study was to compare the performance of FUSION and PIXYL methods to identify treatment failures. Methods. This was a retrospective study of MS patients followed at the Rouen University Hospital. Treatment failure was defined as the occurrence of clinical activity, relapse or disability worsening, or a change in treatment for inefficacy, from the initial MRI up to one year after follow-up MRI. The area under the sensitivity/specificity curve to identify patients with treatment failure (AUC) was the main statistical indicator used to compare the two methods. Results. Treatment failure was observed in 33/83 patients (40%). The distribution of the radiological markers FUSION and PIXYL differed between the groups with and without failure. The PIXYL markers of counts and volumes of new or enlarged lesions were studied. The AUC for the number of new lesions by the two methods did not differ. The optimal sensitivity/specificity thresholds for the number of new lesions were two for FUSION and five for PIXYL rather than the usual benchmark of three new lesions. The combination of markers FUSION and PIXYL in a logistic regression model improved the AUC by 25% (0.836) and doubled the sensitivity (64%), making it possible to identify 33% of additional patients with treatment failure for a conserved specificity (96%). The number of new infratentorial lesions was the most contributing marker of PIXYL. PIXYL segmentation masks reading suggested to remain cautious due to the presence of fake active lesions in 76% of patients. Conclusion. This study was the first clinical evaluation of an automated segmentation method. Classification performances of PIXYL markers synthesized by AUC sensitivity/specificity were not superior to FUSION in identifying patients with treatment failure in the year after follow-up MRI. Any significant gain in sensitivity came at the cost of a loss of specificity incompatible with clinical use. However, the combination of FUSION with PIXYL resulted in a logistic regression model whose performance improved that of FUSION alone. However, caution applies to the conclusions of this exploratory study. In fact, PIXYL segmentation masks reading found many false positives, including contrast artefacts of doubtful significance.
Document type :
Master Thesis
Complete list of metadata

Contributor : Claire Leboeuf Connect in order to contact the contributor
Submitted on : Friday, January 7, 2022 - 2:33:08 PM
Last modification on : Friday, February 18, 2022 - 3:29:41 AM
Long-term archiving on: : Friday, April 8, 2022 - 7:25:30 PM


Mereaux Jean Loup.pdf
Files produced by the author(s)


  • HAL Id : dumas-03516897, version 1


Jean-Loup Méreaux. Comparaison d'une analyse automatique d'IRM cérébrale par PIXYL versus imagerie de fusion dans le suivi de la sclérose en plaques. Médecine humaine et pathologie. 2021. ⟨dumas-03516897⟩



Record views


Files downloads