Deep Learning Based Blood Abnormalities Detection As a Tool for Vexas Syndrome Screening - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Blood Année : 2023

Deep Learning Based Blood Abnormalities Detection As a Tool for Vexas Syndrome Screening

Maxence Bauvais
  • Fonction : Auteur
Pierre Sujobert
  • Fonction : Auteur
Mael Heiblig
  • Fonction : Auteur
Maxime Jullien
  • Fonction : Auteur
Baptiste Le Calvez
  • Fonction : Auteur
Camille Richard
  • Fonction : Auteur
Valentin Le Roc'H
  • Fonction : Auteur
Emmanuelle Rault
  • Fonction : Auteur
Olivier Herault
  • Fonction : Auteur
Pierre Peterlin
  • Fonction : Auteur
Alice Garnier
  • Fonction : Auteur
Patrice Chevallier
  • Fonction : Auteur
Simon Bouzy
  • Fonction : Auteur
Yannick Le Bris
  • Fonction : Auteur
Antoine Neel
  • Fonction : Auteur
Julie Graveleau
  • Fonction : Auteur
Perrine Paul-Gilloteaux
  • Fonction : Auteur
Olivier Kosmider
  • Fonction : Auteur
Marion Eveillard
  • Fonction : Auteur

Résumé

Introduction The VEXAS syndrome (vacuoles, E1 enzyme, X-linked, auto-inflammatory, somatic) described in 2020 caused by mutations of the UBA1 gene, displayed a large pleomorphic array of clinical and biological features. Nevertheless, these criteria do not allow to discriminate VEXAS from other myeloid malignancies on complete blood count + differential, notably due to the absence of peripheral blood characterization of the disease. This study aimed at singling out dysplastic features indicative of VEXAS among peripheral blood (PB) polymorphonuclears (PMN) from VEXAS patients compared to healthy patients and myelodysplastic (MDS) patients. As such task is tedious and subject to operator bias, a multicentric dataset has been used to design a deep learning algorithm for automatic detection of these features, finally tested on an external validation cohort. Patient, material and methods Patients written consents were obtained and five academic centers (number from 1 to 5) were involved in this 3 steps study. Firstly, 3 academic centers (number 1 to 3) enrolled 20 patients distributed as follow: 9 patients with UBA1 mut and 6 patients UBA1 wt with no other genetic mutation as well as 5 samples from MDS patients. A total of 25 PB smears and especially PMN images were gathered and screened for various abnormalities. Two independent morphologists blindly quantified PMN predefined characteristics. This dataset of 2,824 multilabelled PMN was evaluated and tested by a two-sample Wilcoxon Rank Sum test for statistical significance between UBA1 mut and the others. Secondly, in order to automate the detection of these significant features on PMN images, a convolutional neural network (CNN) was trained using a multicentric image dataset gathered from 4 academic centers (number 2 to 5). Patients were selected based on clinical and biological symptoms suggesting VEXAS syndrome, namely fever, skin lesions, chondrites, vasculitis and/or anemia. After evaluation of the UBA1 mutation status, this multicentric cohort was separated into confirmed VEXAS (n = 19) and UBA1 WT (n=20). A total of 6,615 annotated PMN images were collected from DI systems (Cellavision ®, Lund, Sweden) with 1 to 12 PB smears per patient, yielding a 48 UBA1 mut and 33 UBA1 wt smears. All patients were males, the median age was 74 year-old (IQR 67-77) and that of control patients 72 year- old (61-79). Finally, the CNN was evaluated on an external cohort from another academic hospital (center 1), including inflammatory patients, and patients screened for VEXAS syndrome for a total cohort of 15 patients and 1,887 images. Results Four specific abnomalities were observed in PMNs from VEXAS patients as compared with healthy or myelodysplastic controls. The automatic recognition of anomalies was then cast in a multilabel classification task, where each PMN image could be assigned to one or more labels during interference. In order to take into account this specificity, a CNN was trained, composed of different layers of convolutions and pooling (Figure1). The model was set-up to output an independent binary prediction for the presence or absence of each of the four anomalies of interest. Automatic detection of these 4 anomalies by the proposed model yielded area under the curve (AUC) of 0.827; 0.837; 0.927 and 0.947 (Figure 2). Regarding the general performance on the external validation cohort, the CNN achieved a Hamming loss of 0.141, and macro and micro F1 scores of 0.588 and 0.668 respectively. Discussion This study suggests that computer-assisted analysis of PB smears, focusing on suspected VEXAS cases, could provide valuable insights to determine which patients should undergo molecular testing. A deep-learning approach leveraging previously identified peripheral blood indicators and automatic analyzers is thus presented, which can help hematologists orient their suspicion before initiating further analysis.

Domaines

Hématologie

Dates et versions

hal-04456433 , version 1 (14-02-2024)

Identifiants

Citer

Cedric de Almeida Braga, Maxence Bauvais, Pierre Sujobert, Mael Heiblig, Maxime Jullien, et al.. Deep Learning Based Blood Abnormalities Detection As a Tool for Vexas Syndrome Screening. Blood, 2023, 142 (Supplement 1), pp.3648-3648. ⟨10.1182/blood-2023-186874⟩. ⟨hal-04456433⟩
12 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More