Femoral plaque samples were collected during endarterectomy of the femoral artery bifurcation between October 2014 and January 2017 . The severity of LEAD of the patients (N=90) was determined by ankle brachial index (ABI), toe pressure (TP), Fontaine classification of LEAD symptoms, and urgency of the operation. Preoperative magnetic resonance angiography images were used to confirm the severity of the stenosis (Fig. 1).
A and B are preoperative magnetic resonance images of two male patients that present obstructive level of stenosis of femoral artery (blue arrow) caused by the atherosclerotic lesions C and D respectively. Obstruction in these lesions is demonstrated histologically (arrowhead). Lesion C is dominated by NodCa, closely seen in the magnified image (thin arrow) surrounded by fibrin (white asterisk), while lesion D has mainly ShCa (black asterisk) in the magnified image. The trained algorithm (E) and (F) recognizes and calculates area proportion of NodCa (blue color) and ShCa (red color) to the sectioned plaque (green color) area. The calculated area proportion in E is 0.26 for NodCa and 0.10 for ShCa, while in F is 0.06 for NodCa and 0.40 for ShCa.
We compared the area proportion of NodCa and ShCa observed in the plaque tissue to the clinical characteristics listed above. We analyzed endarterectomy samples, which fulfilled two criteria:1) more than 90% stenosis including obstruction, was observed both histologically and by magnetic resonance angiography, and 2) internal elastic lamina for vessel diameter measurement was identifiable in the plaques’ histological section.
Plaques were formalin-fixed, decalcified, and longitudinally sectioned into two halves. Sections were stained with Hematoxylin and eosin stain for histomorphometry analysis. Sample processing and laboratory analyses are described in detail elsewhere .
The patients signed a written informed consent to the study before endarterectomy operation. The study is approved by the Ethics Committee of medicine of the Hospital District of Helsinki and Uusimaa (ASO-project Drno 78/13/03/00/2014) and all methods were performed in accordance with the relevant guidelines and regulations.
Deep learning algorithm training
Hematoxylin and eosin-stained slides of femoral plaques longitudinal sections were digitized with a whole-slide scanner (3D HISTECH Pannoramic 250 Flash III, 3DHistec, Budapest, Hungary) with 20x objective and a pixel size of 0.23µm. The slides were then uploaded to a cloud-based image deep learning platform (Aiforia Create, Aiforia Technologies Oy, Helsinki, Finland).
To quantify each of the predefined calcification categories, two sequential algorithms were developed; the first algorithm, the plaque tissue algorithm recognized and quantified the area of plaque tissue from slide background. The algorithm was set to region context size of 50 µM, the complexity level of Complex and the default specifications. Upon this algorithm, a stratified second algorithm, the calcification algorithm, was built for quantification of the calcification categories, NodCa and ShCa. While developing the algorithms, accuracy was assessed through 1) Verification of each annotation 2) Analysis of the untrained regions and whole section slides. Calcification algorithm was fine adjusted on the following parameters: iterations = 7000, field of view = 100 x, image augmentation range (-10 to 10), aspect ratio=1, maximum sheer=1, luminance range (-30 to 30), contrast range (-30 to 30), maximum white balance change= 10, and noise= 2. The algorithm quantified ShCa and NodCa as the area of every recognized structure of the category in mm2, the collective area of each category, and the area proportion of the calcification category to that of the plaque section tissue area (Fig. 1). To optimize the analysis, visual validation of the final algorithm analysis was conducted, whereafter, tiny lesions that were part of the continuum yet did not contribute to an actual nodule, were excluded. This was done after visually determining the cut-off size limit of the lesion per each slide. The precision of the finalized algorithm was 97.53%, sensitivity was 97.66% and total area error (false positive and negative) was 2.06%.
The maximum width of the vessel diameter in the most stenosed part was measured to assess the vascular remodeling in relation to the area proportion of NodCa and ShCa. This was done using the measurement tool in the Aiforia platform (supplementary figure S1).
Area proportions of NodCa and ShCa was analyzed in relation to patients’ continuous and categorized binomial variables. Continuous data analysis of the patients is presented as the mean (± standard deviation). Data were analyzed for normal distribution by Shapiro-Wilk test. Normally distributed data were analyzed by two-tailed t-test and Pearson correlation test, while non-normally distributed data were analyzed by Mann-Whitney U and Spearman rank analysis. Analysis of covariance was set to assess vessel diameter association with area proportion of NodCa and ShCa along with other confounding factors; age, gender, body mass index (BMI), smoking, hypertension, glomerular filtration rate (GFR) and inflammatory state (high sensitivity C-reactive protein (hs-CRP)). Our data fit the assumption of logistic regression analysis for the association of area proportion of NodCa and ShCa with the clinical parameters of LEAD, adjusted by hypertension, diabetes and dyslipidaemia. P value < 0.05 was considered statistically significant. Data were analyzed using SPSS 25 (Armonk, NY: IBM Corp).
The age was categorized by the median into patients older or younger than 70.5 years. was categorized by the cutoff point of normal and overweight (25 kg/m2) into normal, and overweight or obese (combined). Laboratory measurements were categorized according to the standardized age/gender-relevant reference values that are adopted in the analyzing facility, Helsinki University Hospital Laboratory Services (HUSLAB). Hs-CRP was considered increased for females when levels exceeded 2.5 mg/L and for males when levels exceeded 3 mg/L. GFR was considered impaired if it was less than the following values measured in ml/min /1.73 m2: 77 for patients aged 50-59 years, 69 for patients aged 60-69 and 59 for patients aged 70 years and older. Leukocytosis was labelled for readings higher than 8.2 E9/L, while anaemia was deduced from females’ haemoglobin readings < 117 g/L and from males’ readings < 134 g/L. The categorization of ABI and TP parameters was based on the clinical reference . ABI less than 0.4 and TP less than 30 mmHg were deemed as severe disease indicators. Fontaine class was categorized into two groups, patients with claudication were deemed to have mild symptoms and patients with rest pain, ischaemic ulcer or gangrene were considered to have severe symptoms . Surgical interventions were classified into elective operations, semi-urgent operations were determined if an exceeded intervention was mandated within 4 weeks of their clinical evaluation.