Unbiased Morphometric Assessment of Red Blood Cell Storage Lesion in the Presence of Shear-Induced Stomatocytes

Abstract

Introduction

Red blood cells (RBCs) undergo progressive biochemical and morphological changes during storage, collectively called storage lesion. The quality of red cell concentrates (RCCs) is typically assessed by quantifying hemolysis. An assessment of morphological changes, associated with low quality RBCs, could give an additional indication of the safety and efficacy of the concentrates. The current standard for determining morphological changes is a manual, laborious, and subjectively biased microscopic process that limits the number of cells that can be examined. When using alternative methods like flow cells, flow and shear-induced morphologies affecting especially stomatocyte morphologies must be taken into account. We already established an automated flow morphometric RBC analysis system as an alternative to manual microscopic evaluation. The goal of the present work is to obtain a robust, automated, morphology-related signal (lesion index) quantifying RBC storage lesion in a laminar flow channel under conditions similar to stasis that is not affected by shear-induced reversible morphology changes.

Methods

We use a convolutional neural network (CNN) for high throughput classification of RBCs. We analyzed the morphological changes of 5 RCCs over a period of 12 weeks and classified RBC morphologies, including such that are degradation-induced and reversible. We introduce a lesion index to denote the percentage of irreversible spherical morphologies, known to reduce the post-transfusion survival of erythrocytes. We further addressed shear-induced stomatocyte morphologies in laminar flow and whether these affect CNN-based RBC classification.

Results

Our flow morphometry system achieves a high-resolution classification comprising nine morphological classes with an excellent overall accuracy of 92% and F₁ scores between 84% and 97%. We generate strong evidence that the morphological lesion index can predict the hemolysis level in RCCs during storage. The power of this new classification technique allowed it, for the first time, to detect and measure the lateral concentration gradient of stomatocytes in a conventional flow chamber. Importantly, we show that reversible shear rate-induced morphologies, typical for microfluidic systems, bear no influence on the lesion index.

Conclusion

Flow morphometry combined with evaluation by a CNN allows to reliably assess RBC storage lesion and thus concentrate quality. Additionally, this method reduces the need for complex laboratory procedures.

Introduction

Red blood cell (RBC) transfusion is routinely performed to improve tissue oxygen supply in patients with reduced hemoglobin levels and oxygen-carrying capacity. Typically, blood banks process and store packed RBCs as red cell concentrates (RCCs). During storage, the RBCs undergo progressive biochemical and morphological alterations referred to as storage lesion [1–4]. According to the recommendations of the Council of Europe [5] and the German regulatory guideline [6], the quality of stored RBCs is assessed by measuring the level of hemolysis. To comply with these regulations, the hemolysis level of an RCC is not allowed to exceed 0.8%. However, the hemolysis rate gives only an indication about the already lysed RBCs, which does not indicate the level of deterioration of aged cells known to compromise the post-transfusion survival of erythrocytes [7]. Morphological analysis has more power to predict actual post-transfusion safety and efficacy of stored RBCs [3, 7–10].

Healthy discocytes (Ds) can transform into morphologies of stomatocytes (STs) and echinocytes [11, 12]. This transformation can occur reversibly within seconds if induced by sudden stress [13], but it can also develop irreversibly and slowly due to lesion of the RBCs [14]. At advanced stages of lesion, the RBCs morph into an irreversible spherical shape [15]. The progressively increasing concentrations of such morphologies are a particularly important hallmark of storage lesions [16–18]. Spherical morphologies are rigid cell bodies unable to pass small capillaries. Furthermore, they exert highly compromised oxygen-carrying capacity [19, 20]. Hence, an increasing number of spherical RBCs impair the safety and efficacy of RCCs.

The gold standard for morphological characterization of RBCs is the blood smear test [21], a manual, labor-intensive procedure generating limited statistics of RBCs and prone to subjective bias [18, 22]. As an alternative to blood smear, microfluidic systems with suspended RBCs in laminar flow can provide fully automated morphological diagnosis based on image analysis with large statistics of cells [18, 22–24]. The stronghold of these techniques is their ability to generate large amounts of image data in a short time. This enables extensive data analysis and high signal to noise ratios of extracted signals. We established a flow morphometric analysis system in the past and used it to predict the morphological categories of RBCs within RCCs upon storage [22]. Classification systems for RBCs in laminar flow must consider certain rheological effects, which influence the classification results. Given their biconcave non-spherical shape, RBCs of the same morphology can project different images according to the cell´s random orientation with respect to the optical axis. More importantly, if the shear stress in the flow reaches values in the order of 1 Pa, this can lead to certain reversible changes in morphology [13, 16, 25–28], in particular the generation of shear-induced STs and echinocytes.

In this study, we aim at monitoring specifically those cell morphologies, which characterize the cells in static blood bags. These are essentially the same as the ones in blood smears, stasis, and in studies of flow with low shear stress as in References [23, 29, 30]. However, in order to obtain sufficient statistics in a short time, we generate a moderate flow of the RBC suspension. This does not mean that we have to approach physical conditions like the ones used in References [31, 32] where narrow capillaries and/or high shear stress (>>10 Pa) cause specific cell shapes like croissant or slipper morphologies. Contrary to this scenario, a “quasi-stasis” flow is used in that its shear stress is kept below a threshold of approximately 1 Pa. This is achieved by combining slow flow with sufficiently wide cross-sections (see below). Consequently, we aim to simultaneously detect the degradation-induced morphologies in blood bags on the one hand and the shear-induced morphologies on the other hand. A sufficiently resolved classification with respect to all of these RBC morphologies is therefore needed.

In a previous study, we have used a decision tree algorithm, which due to its limited classification resolution [22, 23, 33] cannot differentiate the ST morphologies. It does therefore not allow to analyze the role of the shear-induced morphologies in the context of measuring the RBC quality. Therefore, we aim at assessing the morphological spectrum by combining the superior classification power of a convolutional neural network (CNN) with the statistical power of in situ microscopy in a laminar flow channel.

CNNs have proven to be very useful in classifying RBCs under static conditions and in flow. They frequently outperform conventional image analysis such as thresholding, edge detection, and mathematical morphology of microscopic images. Their application for RBCs under static conditions [29, 34–36] includes analysis of blood smear samples [37] and extends toward 3D images and holographic microscopy [38, 39]. Static conditions have advantages in, for example, the analysis of RBC shape details [38] or when employing optical microscopic conditions as in routine diagnosis [37]. CNN-guided image analysis is also applied to RBCs in flow, as a step forward from conventional image analysis [22, 23] and employing different flow conditions as, for example, imaging flow cytometry [24], flow chambers with diameters of up to millimeters [40], or microfluidic flow systems with diameters in the micrometer range [30, 31, 41]. As mentioned above, different flow conditions have different applications: flow channels in the µm range with, for example, a cross-section of 8 × 11 μm as in Reference [32] resemble capillaries and allow the detection of flow and shear-induced slipper- and croissant-shape morphologies, addressing deformability as a main characteristic of RBC biological activity. Larger channel cross-sections like the one we use here (1,000 × 100 μm) in the context of “quasi-stasis” (see above) enable assessment of various shear conditions ranging from zero shear in the center of the channel to maximum shear at its walls (compare online suppl. Fig. 1; for all online suppl. material, see https://doi.org/10.1159/000539882). Insofar, and by employing an inexpensive off-the-shelf microfluidic slide, it is possible to assess RBC morphologies similar to stasis and shear conditions causing rolling mode [22, 25] and shear-related morphologies close the chamber walls.

We would like to emphasize that this work is not dedicated to developing our own original CNN architectures. We employ GoogLeNet, a ready-to-use and commercially available CNN provided within the MATLAB environment. The goal of this study is to obtain a robust morphology-related signal (lesion index) that is not affected by shear-induced reversible morphology changes. We show that a suitable lesion index can be obtained by accurate selection of spherical morphologies alone. This marker displays a good correlation with the hemolysis level and thus predicts the quality of RCCs according to official guidelines [5, 6].

Methods

Flow Morphometry Setup

Microscopic and Fluidic System

The experimental setup is based on an in situ microscope as described previously [22, 42, 43]. It has to be emphasized that the in situ microscope generates sharp cell images only within a narrow observational fluid layer whose thickness is optically defined by a focal range of approximately 8 µm. All cells from outside this virtual sampling volume are not imaged or imaged with blur. This has two important consequences: first, the focal range is small as compared to the height of our flow channel (100 μm, see below). This enables the probing of the fluid at different height positions in the channel. Second, the micrographs contain a statistical distribution of cell images with respect to sharpness versus optical blur. Therefore, the image evaluation algorithm has to select the sharp cells for further processing while the blurred cells are being left out. This procedure does not introduce any bias since it is equivalent to the definition of a small observational zone (virtual sample volume) (for further information on the microscopic setup, see the online suppl. data).

We use a microfluidic chamber made from a hydrophobic polymer (ibidi GmbH, Martinsried, Germany) with the dimensions 17.1 × 1.0 × 0.1 mm (L × W × H). A syringe pump moves the RBC suspension through the flow chamber at a flow rate of 0.1 mL/min corresponding to a mean velocity of 1.67 cm/s. The velocity gradient (shear rate) is given by the equation (formula 1: shear rate equation):

$$\text{shear}\text{rate}=\dot{\gamma }\left(\mathrm{h}\right)=\frac{6\bullet F}{B\bullet H^3}\bullet \left|\left(H-2\mathrm{h}\right)\right|,$$

where H = height of the rectangular flow channel’s cross-section; B = width of the rectangular flow channel’s cross-section; h = variable height above the bottom surface; F = flow rate = transported volume over time.

The corresponding shear stress is $\tau =\eta \bullet \dot{\gamma }$, where η is the dynamic viscosity. It is approximately equal to that of water at 20°C, i.e., η = 1.0∙10⁻³ Pa∙s [13, 44]. Within our set of flow-channel parameters, we calculate a shear rate of 760 s⁻¹ at a height h of approximately 12 μm above the bottom and equally at 12 μm below the top wall. The shear stress at these positions is accordingly the following (formula 2: shear stress calculation):

$$\tau =\eta \bullet \dot{\gamma }=0.76\text{Pa}$$

At such values of shear stress in a laminar flow channel, healthy Ds adopt a motion as if they roll on a surface parallel to the channel walls and in the direction of the flow. This phenomenon is termed as rolling mode [22, 25]. We use this distinct feature for specific and sensitive discrimination of healthy discoids. Therefore, we adjust the focus (focal range of appr. 8 µm) of our objective to image selectively in the height position of the rolling mode cells (see online suppl. Figure 1).

Image Processing

The captured images are pre-processed by a custom-made algorithm, which was developed in MATLAB (the MathWorks Inc., Natick, MA, USA). For each detected cell, an individual micrograph (70 × 70 pixels) is generated. Examples of micrographs are shown in Figure 1a (the MATLAB script is available from the authors on request).

Fig. 1.

RBC morphology classes and CNN performance. a Morphologies of RBCs as detected in the experiments described here: in row I, the nine classes (D, T, DS, ST, E1, E2, E3, SE, SP) are represented by typical cell micrographs at 12 μm above channel bottom. This position simplifies the CNN classification due to the flow-related feature “rolling mode” of D, E1, and E2. In row II, we show additional micrographs taken 50 μm above channel bottom, where the rolling mode is not present. The red boxes indicate the pooling for comparison with Reference [22], enabling a reasonable matching with the only three classes of Reference [22] (see online suppl. material). b Confusion matrix (CM), which measures the performance of the CNN with respect to the manual reference of test set.c Class-specific quality indicators of the trained CNN.

Convolutional Neural Network

CNNs are neural networks that are particularly efficient in detecting and classifying specific objects in digital images [45, 46]. We employ GoogLeNet, a CNN architecture provided within MATLAB. This network operates with the special features of inception modules [47, 48] and its neurons are pretrained to classify general objects in images. The specific classification of RBCs requires an additional, specific training of the convolutional kernels, i.e., weights and biases of the neural network. This is done by using a large data set of manually pre-classified erythrocyte micrographs.

Organization of Data for CNN Training and Measurement

For this study, we generated three data sets of images, using our in situ suspension microscope:

• 1.Training data: a set of 10.5 thousand micrographs captured from 12 RCCs, manually classified
• 2.Test data: a set of 1.5 thousand micrographs captured from the same 12 RCCs as the training data, manually classified.
• 3.Measurement data: a set of more than 900 thousand micrographs captured from 5 additional RCCs, not classified.

The micrographs of the training data set were manually classified in nine morphology classes. These pre-classified data were exclusively used for the training of the CNN (for further information on the CNN training, see the online suppl. material). As a performance test after the training, the CNN was applied to classify the micrographs of the test data set. A confusion matrix (CM) was used to validate the CNN performance.

Subsequently, the trained and tested CNN was applied to classify the micrographs of the measurement data set. No manual classification was available for reference. However, a previous study by Sierra et al. [22] had classified the same set of micrographs into three classes by using conventional image processing techniques to extract features and classify the cells based on a decision tree. That classification can be compared to our CNN classification if our nine classes are pooled to match the three classes of the decision tree algorithm.

RBC Sample Preparation

All RCCs used in this study were provided by the German Red Cross Blood Donor Service in Mannheim. No additional ethical vote was required, given that healthy donors signed an informed consent form that specifically informed about the eventual use of the blood product for quality control and research and development purposes.

To generate the sets, training data (consisting of 10,500 micrographs) and test data (consisting of 1,500 micrographs), we monitored a group of twelve RCCs stored in saline adenine glucose mannitol during 5 weeks. Sampling and image capturing was carried out weekly.

To generate the measurement data set (consisting of ∼900,000 micrographs), samples from an additional group of 5 RCCs stored in saline adenine glucose mannitol were taken at week 2, 5, 6, 8, 10, and 12. Given that the shelf-life of RCCs in Germany is 6 weeks [49], this timing schedule ensured the appearance of observable advanced morphological changes upon storage lesion. Before sampling, RCCs were agitated for 20 min by rotating at 0.1 rps (ACR Rotator, Lmb Technologie GmbH, Schwaig, Germany). Hemolysis levels were measured (see online suppl. material) and aliquots for the microfluidic system were prepared as follows: 1.5 μL samples were extracted from RCCs and subsequently diluted in 0.9% NaCl solution at a ratio of 1:1,000.

Statistics

Statistical analysis was performed using GraphPad Prism 9.3.1 (GraphPad Software, San Diego, California, USA). Data representing multiple measurements are depicted as mean values with 95% confidence intervals as error bars.

Simple linear regression was used to calculate the correlation between morphological proportions and hemolysis level. The linear determination coefficient was used as a measure for the goodness of fit.

Results

Examples of CNN-Classified Micrographs and the Classification Accuracy with respect to Manual Sorting and Decision Tree Sorting

Our CNN successfully achieved a morphological resolution corresponding to the following nine RBC classes: Ds, trilobes (Ts), deformed stomatocytes (DSs), stomatocytes (STs), echinocytes1 (Es1), echinocytes2 (Es2), echinocytes3 (Es3), spheroechinocytes, and spherocytes (SPs). Examples of micrographs of each class in different orientations are shown in Figure 1. The sequence E1→E2→E3 of echinocyte morphologies is usually interpreted as a progressive RBC degradation, although it is known that these morphologies occasionally reverse into the “healthy” discoid morphology. The ST classes ST and DS as well as the Ts are known to be reversibly inducible by shear stress [13, 50]. The last stages during degradation comprise the spherical morphologies SE and SP [14, 15, 50], the only classes, which unambiguously represent irreversibly damaged cells.

To evaluate the classification accuracy of the trained CNN, the RBC micrographs of the set test data are used. The results are organized as a CM [51] shown in Figure 1b. This CM serves to compare the CNN classification of individual RBC micrographs with our manual reference classification. As an indicator of the network performance, an overall accuracy can be extracted from the CM. It is defined as follows: accuracy = (number of correctly classified RBCs)/(total number of analyzed RBCs).

From the CM in Figure 1b, we obtain an overall accuracy of 92%. We used F₁ scores to represent a class-specific quality indicator of the CNN [48]. Our nine F₁ scores were as follows: T = 89.9%, DS = 84.1%, ST = 97.1%, D = 96.2%, E1 = 89.6%, E2 = 87.7%, E3 = 89.9%, SE = 87.7%, and SP = 94.0%. The lowest F₁ score (84.1% for DS) is so small because it is based on very few classification events (see numbers in the CM of Figure 1b). To further validate the CNN classification of the micrographs in the measurement data set, we compared it to a low-resolution classification of these same data as carried out by Sierra et al. [22] – with very good results (see online suppl. material).

Shear-Induced ST Morphology in Laminar Flow

In a laminar flow channel, it is expected that positions close to the bottom and top boundaries may cause an accumulation of shear-induced ST and T morphologies. To detect such variances, we tuned the narrow focal range (width of appr. 8 µm) of our optical sample volume over the whole range (100 μm) of height positions h of the flow chamber and correspondingly over a range of shear stresses between 0 Pa and approximately 1 Pa.

In Figure 2, the concentrations of the morphology classes Ts, DSs, and STs are plotted as function of h. These data were taken from one individual RCC in week 1 during storage.

Fig. 2.

Shear rate-induced ST generation close to the walls of the flow channel. a RBC-class concentrations of one individual RCC (stored for 1 week) are plotted as function of position h in the flow channel. h is the height above the channel bottom. The top is at h = H = 100 μm. The measured morphology proportions are plotted as bars against the focal plane position given as height h. The uncertainty of these values can be estimated as follows: through experiments with technical replicates of RCCs, we know that the intra-donor confidence interval (90% confidence) is +/− 3% of absolute values. b The corresponding shear rates are plotted as functions of h.

The distribution of DSs and Ts as functions of the height h in the flow channel shows no significant trends. The corresponding fractions are small but nonzero throughout all positions and shear rates, which agrees with literature data [13, 23]. As these concentrations show no significant dependence on the varying shear stress in the channels cross-section, it can be assumed that in our quasi-stasis condition due to its low shear stress (<1 Pa) they result mainly from cell-intrinsic degradation [52].

The same applies to all other morphology classes (D, E1, E2, E3, SE, SP) since they show no significant concentration gradients over the range of channel heights. Contrary to the behavior of all other classes, the ST distribution exhibits strong maxima close to the chamber walls, where the shear rate is typically 760 s⁻¹, corresponding to a shear stress of 0.76 Pa. The local ST proportion in both these positions is approximately 15%, whereas it is only approximately 5% in the center between the channel walls, where no shear stress occurs.

Lesion Index as a Monitor of Storage Lesion in Individual RCCs

At advanced stages of lesion, RBCs morph into irreversible spherical shapes (SE and SP) [15]. These concentrations do not suffer interference from reversible morphological changes. Therefore, we select these two signals to provide the data for monitoring the RBC quality. To use the information from both signals simultaneously, we defined the lesion index as the sum of the SE and SP concentrations with equal weights.

In a previous study [53], a morphology index (MI) has been defined as a weighted sum of all class concentrations in Reference [54], thereby including the nonspherical morphologies. We also calculated a modified version of that index from our own data (see online suppl. material) so that the MI as another candidate for monitoring cellular lesion can be compared with the lesion index.

The official standard-of-quality assessment of RCCs is based on the measurement of the hemolysis level. To investigate the correspondence of our lesion index with the hemolysis level, we simultaneously collected the hemolysis level and the in situ micrographs from those five RCCs that were used to generate the measurement data (see section RBC sample preparation). Therefore, we collected the values of the hemolysis level of all five RCCs simultaneously with their morphometric concentration data.

For each individual class signal as well as for the MI and the lesion index, we computed the determination coefficients with respect to the hemolysis signal. The mean values were as follows: T R² = 0.109, DS R² = 0.012, ST R² = 0.103, D R² = 0.327, E1 R² = 0.125, E2 R² = 0.028, E3 R² = 0.002, spheroechinocyte R² = 0.514, SP R² = 0.801, MI R² = 0.665, lesion index R² = 0.837.

Among all morphometric signals, the lesion index shows the best correlation to the hemolysis values. This corroborates the notion that only the spherocytic morphologies represent a stage within the progression of lesion damage of RBCs, that is, on a morphological level, close to hemolysis. To illustrate the accordance of the progressively increasing lesion index with the progressively increasing hemolysis, the corresponding measurements for each individual RCC are displayed in Fig. 3a–e.

Fig. 3.

Comparison between lesion index and hemolysis in individual RCCs. a–e For five individual RCCs, the lesion indices are plotted (as bars) against the time. For comparison, the corresponding hemolysis levels are plotted as data points with connecting lines for visual guidance. f The mean values of the lesion index and the hemolysis level, taken over the set of five RCCs, are plotted against the time. Error bars indicate confidence intervals (95% CI). Dotted lines represent a threshold of the lesion index of 11.1%, which we propose as a criterion for the acceptability of the RCCs, like the official hemolysis threshold.

In all cases, the lesion index and the hemolysis level consistently increase over the entire storage time. Nonetheless, the curves of individual RCCs exhibit considerable differences, corresponding to strong inter-donor variances.

To summarize the general trend of the individual observations (Fig. 3a–e), the average values (taken over the set of RCCs) of the lesion index and of the hemolysis level are displayed in Figure 3f as functions of time. The error bars (95% CI) represent the strong inter-donor variance seen between the cases (Fig. 3a–e).

Lesion Index as Predictor of the Hemolysis Level

The above presented results suggest that the lesion index is a suitable predictor variable for hemolysis. To visualize this conjecture, Figure 4 shows a scatter plot of the values of the lesion index from all five RCCs against the synchronous hemolysis levels.

Fig. 4.

Lesion index as hemolysis prediction marker. This plot represents all data points (lesion index/hemolysis level) obtained during the 12 weeks from the five RCCs corresponding to measurement data. Dotted lines show the threshold values for RBC quality (lesion index = 11.1% and hemolysis level = 0.8%).

We obtain a good correlation (R² = 0.84) between hemolysis level and lesion index, representing the whole data of five RCCs from different donors. Within this set of data, a threshold of 11.1% for the lesion index corresponds to the critical hemolysis level of 0.8%.

As compared to the lesion index, the concentration signals of the classes D and SE exhibit considerably less correlation with hemolysis, and consequently they are less useful for predicting the hemolysis level (see online suppl. Fig. 2). Incidentally, the same applies to the SP subclass SP; this class alone (R² = 0.80) does not perform as well as the lesion index, which comprises not only SP but also SE. Similarly, the correlation of hemolysis with MI (R² = 0.67) is only moderate. This is understandable since the definition of the MI includes concentration signals from morphology classes with more signal noise and weaker correlations to hemolysis.

Discussion

The aim of this study was to obtain a robust, automated, morphology-related signal (lesion index) quantifying RBC storage lesion in a laminar flow channel under conditions similar to stasis robust against shear-induced reversible morphology changes. We have shown that flow morphometry [22, 55] of RBCs using a trained CNN achieves a much more resolved classification as compared to our previous work based on a decision tree algorithm. The CNN-based system discriminates RBCs in nine morphological classes with an excellent overall accuracy of 92% and F₁ scores between 84% and 97%.

These nine classes each represent reversible or irreversible stages in the lifetime of RBCs. The power of this new classification technique allowed it for the first time to detect and measure the lateral concentration gradient of STs in a conventional flow chamber. Our results show, however, that a lesion index that is based distinctively on spherical morphology classes avoids any errors due to the concentration gradients of the ST class. We demonstrate a good correlation between hemolysis rate and the flow morphometry-based lesion index. Accordingly, we propose the latter as a parameter to predict the hemolysis level and to estimate the quality of stored RBCs with respect to post-transfusion survival.

Shear-Induced STs in Flow Chambers

Morphological transients of erythrocytes may not necessarily relate to biochemical alterations and therefore might lead to errors if interpreted as a quality marker. Flow and the concomitant shear rate have been shown to affect RBC morphologies. Especially the formation of ST morphologies (T, DS, and ST) depends on shear stress [13, 16, 27, 28]. In fact, our results reveal for the first time that the proportion of STs increases close to the boundaries of the laminar flow channel. Since this position coincides with the maximal shear stress, we conclude that close to the channel surfaces the larger part of the observed STs consists of formerly discoid and healthy cells, only temporarily deformed by shear stress. This suggests a general conjecture with respect to any morphological RBC characterization based on laminar flow channels, including imaging flow cytometry. If the corresponding measurement setup generates local shear stress above approximately 0.7 Pa, then shear-induced ST morphologies occur at that position [56]. This effect complicates studies aimed at general morphological information, for example, in relation to the stomatocyte-discocyte-echinocyte sequence (see below). However, the shear-induced effects are unlikely to interfere with the concentrations of the irreversible spherical morphologies SE and SP since these are caused by irreversible biochemical alterations. As for the lesion index, its value is built exclusively on the spherical morphologies and hence it should not be influenced by shear. In agreement with this conjecture, we observe that the lesion index does not exhibit the same random fluctuations as the other concentration signals including the MI.

Prediction of Hemolysis Level by Using the Lesion Index

The high-resolution classification performance of the CNN enables the monitoring of characteristic morphological transients caused either by lesion or by exterior stress. In view of monitoring the RCCs quality, we defined a lesion index as the sum of relative proportions of spherical morphology concentrations. In individual RCCs, the values of the lesion index correlate well with hemolysis levels taken simultaneously over storage time (see Figure 4). Our results generate strong evidence that the morphological lesion index can indeed predict the hemolysis level in RCCs during storage. This is important in that the hemolysis level is currently the only biochemical quality parameter in several international guidelines, such as the European guideline for RBC products [5]. If the hemolysis level exceeds the threshold of 0.8%, then that RCC must be discarded. Within our set of data, the 0.8% threshold of the hemolysis level corresponds to a preliminary threshold of 11.1% of the lesion index.

Of course, our numerical value of the lesion threshold is only a preliminary estimate because its statistical basis is limited. To apply a threshold value of this type as a compliance criterion for RCCs, the used flow morphometry method including the trained CNN needs to be standardized and a final numerical threshold value must be determined with appropriate statistics.

As opposed to the spherical morphologies, all other individual RBC classes, including Ds, are subject to reversible transitions among each other. These are related to the so-called stomatocyte-discocyte-echinocyte sequence [11], which includes interactions with shear rates and other temporary stress factors [13, 53, 57]. Influence factors of this kind cause lesser signal to noise ratios of the T, DS, D, ST, E1, E2, and E3 signals and of the MI as well. The correlation of each of these signals with respect to the hemolysis level is therefore weaker than that of the lesion index.

The hemolysis measurement is a marker that indicates the ratio of lysed cells to functional cells present in an RCC. The lesion index, in contrast, reflects a stage earlier in the degradation process of the cells, namely, when they are not yet disintegrated. It thus provides potentially important additional information about the quality of the living cell fraction. The lesion index monitors the appearance of cells with spherical morphology. These will be physiologically eliminated after transfusion, and therefore the lesion index is a plausible parameter to predict a lower limit of post-transfusion survival of blood cells [7, 10]. According to the requirements of the US Food and Drug Administration (FDA), a post-transfusion survival rate of blood cells of at least 80% is targeted.

Consequently, the results of this study suggest that the lesion index can indeed provide a suitable quality criterion for RCCs in clinical routine. Therefore, future studies could assess the predictive value of the lesion index for RCC quality in critically ill patients in terms of transfusability and clinical outcome [58–60]. Also, future studies using our setup could help in elucidating which of the transient morphology changes are reversible after transfusion.

Furthermore, the analysis of large numbers of objects at sub-micrometer resolution in laminar flow could facilitate assessment of subhemolytic shear stress [61–63]. Indicators of such sublethal stress are, among others, cell fragmentation of subpopulations, membrane blebbing, and morphological shape change at stasis [64] – effects that could be visible in a statistically sound way in the system presented here.

Conclusion

Robust classification of erythrocytes can be achieved using a reagent-free method, without extensive microscopic or biochemical analysis equipment, employing a standard laboratory flow chamber with laminar flow combined with in situ microscopy and a trained CNN. A concentration profile is obtained over a range of nine morphology classes, thereby providing a fingerprint of the RBC quality.

Our CNN-based flow morphometry reveals for the first time that the concentration of ST-type erythrocytes rises sharply in the proximity of flow chamber walls where the local shear stress reaches values of typically 0.7 Pa. It is not unusual that local shear stress of this order of magnitude occurs in flow channel setups (including imaging flow cytometry), where the cross-sectional dimensions are an order of magnitude larger than the RBCs themselves. In these cases, a local accumulation of STs should be considered.

By adjusting the position of the focal plane within the laminar flow chamber, it is possible to analyze cells under different, yet defined, shear conditions. The core element of the system, the flow chamber, is off-the-shelf, as are objective, camera and pump; the CNN is available upon request from the authors. Therefore, this setup might constitute an inexpensive and easy-to-use system not only for quality control but also as an alternative to, for example, cone-and-plate rheometers, or assessing the flow regime in larger blood vessels. Additionally, it might be used in the study of sublethal stress.

The concentration of spherical morphologies can be used to define a lesion index as a variable monitoring the RBC quality. The lesion index is shown to be a robust indicator of RBC quality in the sense that it is independent of influencing factors such as shear stress and reversible transitions between different morphologies. It is independent of the focal plane position in the flow chamber despite the concentration gradients of STs.

By using the lesion index as a prediction variable, the compliance of an RCC with respect to the official hemolysis limit can thus be tested without having to perform the hemolysis measurement itself. Given that spherical RBCs are among the cells that will be physiologically removed from the blood circulation after transfusion, the lesion index informs directly about that fraction of the RCC. We suggest the lesion index determined by flow morphometry for routine quality assessment of RCCs: it predicts the hemolysis level, and its numerical value represents a lower threshold for the expected post-transfusion survival of RBCs.

Acknowledgments

We acknowledge the valuable contribution of the staff of German Red Cross Blood Service, especially in the production, issue department, and quality control laboratory. For the publication fee, we acknowledge financial support from Heidelberg University. This work is based on the dissertation of Clemens Boecker, University of Heidelberg, 2023 [65].

Statement of Ethics

This study protocol was reviewed and the need for approval was waived by the Mannheim Ethics Committee II, Medical Faculty Mannheim, Heidelberg University; all donors signed informed consent that blood can be used for research and development purposes.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

Stipends to C. Boecker have been provided by the Albert-and-Anneliese-Konanz Foundation, Mannheim University of Applied Sciences, and the Karl-Völker-Foundation, Mannheim University of Applied Sciences.

Author Contributions

All authors provided substantial contributions to the conception or design of the work, or the acquisition, analysis, or interpretation of data for the work, drafting the work or revising it critically for important intellectual content, final approval of the version to be published, and agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Funding Statement

Stipends to C. Boecker have been provided by the Albert-and-Anneliese-Konanz Foundation, Mannheim University of Applied Sciences, and the Karl-Völker-Foundation, Mannheim University of Applied Sciences.

Data Availability Statement

All data generated or analyzed during this study are included in this article and its online supplementary material. Further inquiries can be directed to the corresponding authors.