Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia

Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia View ORCID Profile Jan-Niklas Eckardt , Ishan Srivastava , Freya Schulze , Susann Winter , Tim Schmittmann , Sebastian Riechert , Martin M. K. Schneider , Lukas Reichel , Miriam Eva Helena Gediga , Katja Sockel , Anas Shekh Sulaiman , Christoph Röllig , Frank Kroschinsky , Anne-Marie Asemissen , Christian Pohlkamp , Torsten Haferlach , Martin Bornhäuser , Karsten Wendt , Jan Moritz Middeke doi: https://doi.org/10.1101/2025.01.27.25321165 Jan-Niklas Eckardt 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany 2 Else Kröner Fresenius Center for Digital Health, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jan-Niklas Eckardt For correspondence: jan-niklas.eckardt{at}uniklinikum-dresden.de Ishan Srivastava 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany 2 Else Kröner Fresenius Center for Digital Health, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Freya Schulze 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Susann Winter 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Tim Schmittmann 3 Institute of Software and Multimedia Technology, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Sebastian Riechert 2 Else Kröner Fresenius Center for Digital Health, TUD Dresden University of Technology , Dresden, Germany 3 Institute of Software and Multimedia Technology, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Martin M. K. Schneider 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Lukas Reichel 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Miriam Eva Helena Gediga 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Katja Sockel 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Anas Shekh Sulaiman 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Christoph Röllig 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Frank Kroschinsky 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Anne-Marie Asemissen 4 Department of Hematology, Oncology and Bone Marrow Transplantation with Section of Pneumology, University Medical Center Hamburg-Eppendorf , Hamburg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Christian Pohlkamp 5 Munich Leukemia Laboratory , Munich, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Torsten Haferlach 5 Munich Leukemia Laboratory , Munich, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Martin Bornhäuser 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany 6 German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ) , Heidelberg, Germany 7 National Center for Tumor Diseases Dresden (NCT/UCC) , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Karsten Wendt 2 Else Kröner Fresenius Center for Digital Health, TUD Dresden University of Technology , Dresden, Germany 3 Institute of Software and Multimedia Technology, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jan Moritz Middeke 1 Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology , Dresden, Germany 2 Else Kröner Fresenius Center for Digital Health, TUD Dresden University of Technology , Dresden, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Evaluation of bone marrow morphology by experienced hematologists is key in the diagnosis of myeloid neoplasms, especially to detect subtle signs of dysplasia in myelodysplastic neoplasms (MDS). The majority of recently introduced deep learning (DL) models in cytomorphology rely heavily on manually drafted cell-level labels, a time-consuming, laborious process that is prone to substantial inter-observer variability, thereby representing a substantial bottleneck in model development. Instead, we used robust image-level labels for end-to-end DL and trained several state-of-the-art computer vision models on bone marrow smears of 463 patients with MDS, 1301 patients with acute myeloid leukemia (AML), and 236 bone marrow donors. For the binary classifications of MDS vs. donors and MDS vs. AML, we obtained an area-under-the-receiver-operating-characteristic (ROCAUC) of 0.9708 and 0.9945, respectively, in our internal test sets. Results were confirmed in an external validation cohort of 50 MDS patients with corresponding ROCAUC of 0.9823 and 0.98552, respectively. Explainability via occlusion sensitivity mapping showed high network attention on cell nuclei not solely of dysplastic cells. We not only provide a highly accurate model to detect MDS from bone marrow smears, but also underline the capabilities of end-to-end learning to solve the bottleneck of time-consuming cell-level labeling. Introduction Myelodysplastic neoplasms (MDS) encompass clonal myeloid malignancies that are characterized by ineffective hematopoiesis, cytopenia, myelodysplasia, and recurrent genetic events. 1 The incidence of MDS appears to be underestimated and incidence rates increase dramatically over the age of 70 years (up to an estimated 75:100,000 cases), representing a substantial societal burden in an aging population. 2 – 4 Although genetic findings are becoming increasingly important according to the new WHO 2022 classification, accurate cytomorphologic evaluation of the bone marrow remains crucial for the initial diagnosis, response assessment, and detection of disease transformation to acute myeloid leukemia (AML). 5 While counting myeloblasts is rather straightforward, signs of dysplasia are more subtle and their accurate identification requires experienced investigators. Still, detection is often challenging and prone to inter-observer variability, even for seasoned morphologists 6 – 8 , and shows discrepancies between site and central review. 9 In general, cytomorphologic evaluation of bone marrow aspirates in hematology remains essentially unchanged over the last decades, as both preparation and evaluation are performed manually, rendering the entire process time- and cost-intensive, as well as dependent on the experience and subjective judgement of the observer. 10 – 12 With the advent of deep learning (DL) systems for computer vision 13 , a multitude of applications in the healthcare sector have been identified where DL is applied in image-based diagnostics. 14 , 15 Convolutional neural nets (CNN), which consist of multiple artificial neurons that are interconnected via convoluted deep layers, are commonly used for computer vision tasks. 16 In cytomorphology, recent studies have utilized neural networks in order to correctly classify peripheral blood and bone marrow cells based on their respective morphology 17 – 28 , as well as to accurately identify myeloid malignancies. 29 , 30 In this study, we used an end-to-end DL system to accurately differentiate between MDS, AML, and healthy donor bone marrow samples based on image-level labels, without the need for manually labeling cells or dysplastic morphologies. Methods Data sets We identified 463 MDS patients that have been previously diagnosed and treated at the University Hospital Dresden, Germany. The first control group comprised 1301 AML patients that had been diagnosed and treated under the auspices of the multicenter German Study Alliance Leukemia (SAL) within the following previously reported multicenter trials: AML96 31 [ NCT00180115 ], AML2003 32 [ NCT00180102 ], AML60+ 33 [NCT 00180167], and SORAML 34 [ NCT00893373 ]. Patients were eligible upon diagnosis of MDS or AML according to the revised WHO/ICC criteria 5 , 35 age ≥18 years, and available biomaterial at initial diagnosis including bone marrow smears. The second control group consisted of 236 bone marrow samples from healthy bone marrow donors who underwent allogeneic bone marrow donation at our center as previously reported. 36 An additional external validation cohort was obtained from the Munich Leukemia Laboratory (MLL), Munich, Germany, consisting of 50 patients with diagnosed MDS according to the above-mentioned eligibility criteria. Prior to analysis, written informed consent was obtained from all patients and donors according to the revised Declaration of Helsinki. 37 All studies were approved by the Institutional Review Board of the TUD Dresden University of Technology (EK 98032010 and EK 289112008). Image digitization Bone marrow smears (BMS) were prepared from anticoagulated bone marrow according to WHO guidelines. 38 Staining of MDS, AML, and donor BMS was performed with the May-Grünwald-Giemsa method. 11 Image-level labels were derived from case-level diagnostics, including cytomorphology, histology, cytogenetics and molecular genetics, previously documented for each case during routine diagnostics or as part of the respective clinical trial. Using a Pannoramic 250 FLASH III (3DHISTECH), we obtained high-resolution whole slide images. For every AML patient and bone marrow donor, one image (50x magnification) per whole slide image was obtained using SlideViewer (3DHISTECH). We assumed that subtle signs of dysplasia would not be fully captured in one field of view alone. Therefore, for each MDS patient, we obtained four pictures (50x magnification) of different areas of interest in the BMS. Deep learning End-to-end image-level prediction on bone marrow slides We extended our previously described DL pipeline 29 , 30 for binary image-level predictions for the delineation of MDS, AML, and healthy controls. Based on case-level diagnosis, images were labeled with either “MDS”, “AML”, or “healthy donor”. Importantly, no cell-level manual labeling was performed. The pipeline was adapted to evaluate cases in a binary fashion, i.e. MDS vs. AML and MDS vs. healthy donors. Potentially, imbalanced training data can bias a classifier towards the predominant class. Considering the imbalances between the data sets (n=463 samples for MDS with 4 images per patient, resulting in 1852 MDS images in total; n=1301 samples for AML with 1 image per patient; n=236 samples per donor with 1 image per donor), we used image augmentation techniques, such as random sized cropping, color shifting and linear transformations, to balance the data sets for each binary classification task. For all binary classifications, a 5-fold internal cross-validation was used, i.e. a train-test-split of 80:20. Cases that were used for model training were strictly separated from cases that were used for testing. In DL, determination of an optimal model cannot be done a priori, but rather has to be evaluated given the specific use case, data set, and model architecture. Hence, we evaluated six recently introduced DL architectures for computer vision including ResNet-18/34/50/101/152 39 , ResNeXt-50_32×4d/101_32×8d 40 , Wide-ResNet-50/101 41 , DenseNet-121/161/169/201 42 , ShuffleNet v2_x0_5/v2_x1_02 43 , and SqueezeNet v1.1 44 . All DL models were pre-trained on ImageNet data. 45 The final architecture for each model was determined using automated hyperparameter optimization with the Optuna framework. 46 DL models were implemented in Python using the PyTorch framework. Computations were performed using the high-performance computing (HPC) cluster of the TUD Dresden University of Technology. Performance evaluation Recall ( syn .: sensitivity), precision ( syn .: positive predictive value), and accuracy were used to evaluate classification performances. Recall is defined as the fraction of all positive predictions among all relevant events and precision is defined as the fraction of true positives among all positive predictions. Further, the area-under-the-curve (AUC) was determined for the receiver-operating-characteristic (ROC). All metrics are reported for each binary classification for the internal test sets as well as for the external validation cohort with 95% confidence intervals. Explainability of classifications via occlusion sensitivity maps To highlight network attention and thereby identify morphological cues the network used to delineate MDS, AML, and healthy donors, we used occlusion sensitivity maps (OSM). In OSM, random image areas are iteratively blocked from view of the CNN and classification performance is measured. If the blocked image area is highly relevant for accurate image-level classification, model performance will drop accordingly. This process is repeated for the entire image. Thus, image areas that are crucial for accurate predictions are highlighted so that morphologies that prompt the CNN classifier to predict a label can be evaluated and interpreted. Results End-to-end deep learning accurately delineates MDS from AML and healthy controls Baseline characteristics of the MDS patient cohort are shown in Table 1 . We evaluated six different neural network architectures 39 – 44 for binary classification tasks iteratively. For the distinction between MDS and healthy donors, we found Densenet-201 42 to provide the highest classification performance, with an accuracy of 0.97791 and a corresponding ROCAUC of 0.9708 ( Table 2 ; Figure 1A ). With respect to delineating MDS from AML, the best results were obtained using the Squeezenet 44 architecture, resulting in an accuracy of 0.98072 and a ROCAUC of 0.9945 ( Table 2 ; Figure 1B ). Detailed information on metrics and 95% confidence intervals of the best performing models for each use-case is provided in Table 2 . Individual model training on the HPC system for MDS vs. healthy donors and MDS vs. AML took 20 hours each. An external validation set encompassing 50 MDS patients was obtained from the Munich Leukemia Laboratory (MLL). Using our pre-trained models, we achieved an accuracy of 0.9972 with a corresponding ROCAUC of 0.9823 in distinguishing external MDS samples from healthy controls ( Table 3 , Figure 2A ). With respect to delineating external MDS samples from AML, an accuracy of 0.92104 was achieved with a ROCAUC of 0.98552 ( Table 3 , Figure 2B ). Download figure Open in new tab Figure 1: Performance of deep learning models for binary classifications delineating MDS, AML, and healthy donors. The receiver-operating characteristic (ROC) with the corresponding area-under-the-curve (AUC) is depicted for the best performing models for each classification task. For MDS vs. healthy donors, best results were achieved with Densenet-201 (A). For MDS vs. AML, best results were achieved with Squeezenet (B). Internal cross-validation was performed with an 80:20 split. Individual run performance (Fold 1-5; graphs in light blue, orange, green, red, and purple) as well as aggregate macro average performance (graph in dark blue) are reported. Only testing results are reported. Download figure Open in new tab Figure 2: External validation of deep learning models for binary classifications delineating MDS, AML, and healthy donors. The receiver-operating characteristic (ROC) with the corresponding area-under-the-curve (AUC) is depicted for the best performing models for the binary classifications MDS (MLL) vs. healthy donors (A) and MDS (MLL) vs. AML (B). Individual run performance (Fold 1-5; graphs in light blue, orange, green, red, and purple) as well as aggregate macro average performance (graph in dark blue) are reported. View this table: View inline View popup Download powerpoint Table 1. MDS patient characteristics View this table: View inline View popup Download powerpoint Table 2. Test set performance for binary image-level classifications View this table: View inline View popup Download powerpoint Table 3. Model performance on external validation set Explainable predictions via occlusion sensitivity maps In order to make results interpretable to cytomorphologists, we used OSM that iteratively blocks image areas from neural network evaluation and thus highlights (in red) image areas that are of high importance for accurate predictions. In a proof-of-concept fashion, we found OSM to be cell-specific, indicating that network attention is being focused on cells, rather than background or smudge. Network attention was focused on cells in granulopoiesis and erythropoiesis and on megakaryocytes ( Figure 3 ). Interestingly, neural networks focused not only on signs of dysplasia, but also on cells we deemed morphologically inconspicuous. High attention was given to defined signs of dysplasia involving altered nuclear morphology such as chromatin clumping, dysfunctional segmentation, or double nuclei. However, at times, high network attention was also given to cells with no apparent dysplasia as per conventional definition, 11 , 47 while network attention in these cells was also mainly focused on the nucleus, sometimes including the perinuclear zone. This indicates more intricate and subtle morphological alterations unquantifiable by human observers. However, other signs of dysplasia, such as hypogranulation, were disregarded by our model. This could possibly be either due to confidence saturation - meaning the model found enough reasons in a given field of view to confidently predict MDS without paying attention to all apparent signs of dysplasia (defined or not) - or low-ranking signs of dysplasia that were not learned in the training process due to their limited weight in making accurate predictions. Download figure Open in new tab Figure 3: Occlusion Sensitivity Mapping (OSM) highlights network attention for explainable output interpretation. OSM iteratively blocks image areas from being evaluated by the deep learning network. If an image area is highly important for classification, the network’s performance will thus drop substantially in the given iteration. Image areas that are of high importance for correct classification can thereby be highlighted (high attention shown in red). A standard field of view of bone marrow smears from MDS patients is shown in A, C, and E. The corresponding OSM is displayed in B, D, and F, respectively. First, in a proof-of-concept fashion, the network focuses its attention on cells and specifically on nuclei. It does not consider background, noise or smudge as important for classification. Second, high attention is directed at erythropoietic and granulopoietic cells as well as megakaryocytes. Discussion Using end-to-end DL, we developed a software framework to distinguish between MDS, AML, and healthy controls with very high accuracy based on BMS from 2000 individual patients and bone marrow donors. Importantly, we have demonstrated that information abstraction even in MDS with often subtle morphologies is feasible using end-to-end learning, in contrast to recent studies in hematology that primarily rely on the generation of cell-level labels. 17 – 28 , 48 , 49 Using the latter approach, a bottom-up system has to be devised where first thousands (usually hundreds of thousands) of labels are required to build a robust classifier, and second individual cell-level predictions have to be aggregated to generate a diagnosis-level prediction. Apart from being obviously time-consuming and cost-ineffective, the generation of cell-level labels, i.e. the ground truth many classifiers in hematology currently are based on, is flawed due to substantial classification biases. For instance, Sasada et al. 7 evaluated divergence in cell classifications on 100,000 hematopoietic cells of 499 MDS patients. Up to eleven experienced observers evaluated each cell image, however, only 55.6% of classifications were found to match and especially low classification overlap was reported for dysplastic morphologies like hypo-granularity and Pseudo-Pelger-Huët anomaly. 7 This bottleneck and pitfall of cell-level labeling can essentially be bypassed by an end-to-end approach, such as ours. Our ground truth labels are not derived from subjective observer judgment. Instead, they are established by routine diagnostics, including cytomorphology, histology, flow cytometry, cytogenetics, molecular genetics and clinical examination, which provide image-level ground truth labels that are much more robust. With respect to explainability, DL is often referred to as a ‘black box’. 50 The often elusive decision-making of neural networks substantially hampers interpretability and thus acceptance of DL models in such high-risk applications as cancer diagnostics. Using OSM (among other methods of explainability 51 ) not only enables internal proof-of-concept, but also provides additional information to the human observer, as novel features that are important for prediction can be investigated that otherwise would elude the human eye. Interestingly, our classifiers showed high attention for nuclei not only of dysplastic cells, but also for cells that we did not deem to be morphologically suspicious for dysplasia. Potentially, this alludes to a digital biomarker in MDS distinct from classical signs of dysplasia. Future work will focus on correlating attention maps with genetic alterations and/or gene expression in MDS. While certain molecular alterations have already been linked to certain morphologies, such as mutated SF3B1 in MDS with ringsideroblasts 52 , 53 , CNNs can potentially be used to identify novel gene-morphology links. For instance, Brück et al. 54 used CNNs on MDS bone marrow core biopsies to predict mutations of TET2 , spliceosome genes and monosomy 7. Further, Nagata et al. 55 previously demonstrated a link between MDS morphology (assessed by pathologists) and genomic profiles. This suggests a starting point for CNNs to link gene alterations with specific morphologies and may potentially lead to an image-based predictor of genetic profiles and consequentially patient risk and outcome. Our study is limited by several factors. As is the case for most recent studies of computer vision in (hemato-)pathology, our analysis is based on retrospective data. While external validation confirmed high classification accuracy, prospective validation is still warranted. In our study, we differentiated only between AML, MDS, and healthy bone marrow donors in a binary way. Still, some dysplastic morphologies can also be present to a certain degree in non-malignant disorders 56 such as congenital syndromes 57 , nutritional deficiencies 58 , 59 , infectious disease 60 , and drug- or toxin-mediated bone marrow damage 61 , 62 . To increase routine applicability of our DL framework, future work will also focus on acquiring image data from reactive and non-neoplastic specimen exhibiting bone marrow dysplasia in order to make our classifier more versatile and applicable in clinical routine. In summary, we have developed a DL framework trained on patient and donor samples, achieving high accuracies in our internal test set and external validation set in distinguishing between MDS, AML, and healthy bone marrow donors. Data Availability All data produced in the present study are available upon reasonable request to the authors. Author Contributions J-NE and JMM conceptualized the project. J-NE, FS, MEVG, KS, ASS, CR, UP, CP, TH, MB, and JMM provided patient samples. J-NE, FS, MS, LR, and MEHG acquired BMS images. IS, TS, SR, and KW developed computer vision models. All authors analyzed and interpreted the data. J-NE wrote the initial draft. All authors approved the final version of the manuscript and agreed to be accountable for all aspects of the work. Competing Interests J-NE, TS, SR, and JMM are co-owners of Cancilico. Acknowledgements The authors are grateful to the Centre for Information Services and High-Performance Computing of the TUD Dresden University of Technology for providing its facilities for training and execution of deep learning models. This study was funded in part by Novartis Oncology. The funder had no role in conceptualization, design, data collection, analysis, decision to publish, or preparation of the manuscript. Footnotes Competing Interests: The authors declare no competing interests. List of Abbreviations AML acute myeloid leukemia BMS bone marrow smear(s) CNN convolutional neural net DL deep learning MDS myelodysplastic neoplasm(s) References ↵ Cazzola M . Myelodysplastic Syndromes . N Engl J Med 2020 ; 383 : 1358 – 1374 . OpenUrl CrossRef PubMed ↵ Greenberg PL , Attar E , Bennett JM , Bloomfield CD , Borate U , De Castro CM et al. Myelodysplastic syndromes: clinical practice guidelines in oncology . J Natl Compr Canc Netw 2013 ; 11 : 838 – 874 . OpenUrl Abstract / FREE Full Text Neukirchen J , Schoonen WM , Strupp C , Gattermann N , Aul C , Haas R et al. Incidence and prevalence of myelodysplastic syndromes: Data from the Düsseldorf MDS-registry . Leukemia Research 2011 ; 35 : 1591 – 1596 . OpenUrl CrossRef PubMed ↵ Cogle CR . Incidence and Burden of the Myelodysplastic Syndromes . Curr Hematol Malig Rep 2015 ; 10 : 272 – 281 . OpenUrl PubMed ↵ Khoury JD , Solary E , Abla O , Akkari Y , Alaggio R , Apperley JF et al. The 5th edition of the World Health Organization Classification of Haematolymphoid Tumours: Myeloid and Histiocytic/Dendritic Neoplasms . Leukemia 2022 ; 36 : 1703 – 1719 . OpenUrl CrossRef PubMed ↵ Goasguen JE , Bennett JM , Bain BJ , Brunning R , Vallespi M-T , Tomonaga M et al. Dyserythropoiesis in the diagnosis of the myelodysplastic syndromes and other myeloid neoplasms: problem areas . Br J Haematol 2018 ; 182 : 526 – 533 . OpenUrl PubMed ↵ Sasada K , Yamamoto N , Masuda H , Tanaka Y , Ishihara A , Takamatsu Y et al. Inter-observer variance and the need for standardization in the morphological classification of myelodysplastic syndrome . Leuk Res 2018 ; 69 : 54 – 59 . OpenUrl PubMed ↵ Valent P , Orazi A , Steensma DP , Ebert BL , Haase D , Malcovati L et al. Proposed minimal diagnostic criteria for myelodysplastic syndromes (MDS) and potential pre-MDS conditions . Oncotarget 2017 ; 8 : 73483 – 73500 . OpenUrl CrossRef PubMed ↵ Zhang L , Stablein DM , Epling-Burnette P , Harrington AM , Moscinski LC , Kroft S et al. Diagnosis of Myelodysplastic Syndromes and Related Conditions: Rates of Discordance between Local and Central Review in the NHLBI MDS Natural History Study . Blood 2018 ; 132 : 4370 . OpenUrl CrossRef ↵ Lee S-H , Erber WN , Porwit A , Tomonaga M , Peterson LC , International Council for Standardization In Hematology. ICSH guidelines for the standardization of bone marrow specimens and reports . Int J Lab Hematol 2008 ; 30 : 349 – 364 . OpenUrl CrossRef PubMed ↵ Bain BJ , Clark DM , Wilkins BS . Bone Marrow Pathology . John Wiley & Sons , 2019 . ↵ Bain BJ , Bailey K . Pitfalls in obtaining and interpreting bone marrow aspirates: to err is human . J Clin Pathol 2011 ; 64 : 373 – 379 . OpenUrl Abstract / FREE Full Text ↵ Krizhevsky A , Sutskever I , Hinton GE. ImageNet classification with deep convolutional neural networks . In: Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1 . Curran Associates Inc. : Red Hook, NY, USA , 2012 , pp 1097 – 1105 . ↵ Chan H-P , Samala RK , Hadjiiski LM , Zhou C . Deep Learning in Medical Image Analysis . Adv Exp Med Biol 2020 ; 1213 : 3 – 21 . OpenUrl CrossRef PubMed ↵ Walter W , Haferlach C , Nadarajah N , Schmidts I , Kühn C , Kern W et al. How artificial intelligence might disrupt diagnostics in hematology in the near future . Oncogene 2021 ; 40 : 4271 – 4280 . OpenUrl PubMed ↵ Dhillon A , Verma GK . Convolutional neural network: a review of models, methodologies and applications to object detection . Prog Artif Intell 2020 ; 9 : 85 – 112 . OpenUrl ↵ Matek C , Krappe S , Münzenmayer C , Haferlach T , Marr C . Highly accurate differentiation of bone marrow cell morphologies using deep neural networks on a large image data set . Blood 2021 ; 138 : 1917 – 1927 . OpenUrl CrossRef PubMed Choi JW , Ku Y , Yoo BW , Kim J-A , Lee DS , Chai YJ et al. White blood cell differential count of maturation stages in bone marrow smear using dual-stage convolutional neural networks . PLoS One 2017 ; 12 : e0189259 . OpenUrl CrossRef PubMed Kainz P , Burgsteiner H , Asslaber M , Ahammer H . Training echo state networks for rotation-invariant bone marrow cell classification . Neural Comput Appl 2017 ; 28 : 1277 – 1292 . OpenUrl PubMed Matek C , Schwarz S , Spiekermann K , Marr C . Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks . Nat Mach Intell 2019 ; 1 : 538 – 544 . OpenUrl Putzu L , Caocci G , Di Ruberto C . Leucocyte classification for leukaemia detection using image processing techniques . Artif Intell Med 2014 ; 62 : 179 – 191 . OpenUrl PubMed Rodellar J , Alférez S , Acevedo A , Molina A , Merino A . Image processing and machine learning in the morphological analysis of blood cells . Int J Lab Hematol 2018 ; 40 Suppl 1 : 46 – 53 . OpenUrl CrossRef PubMed Acevedo A , Merino A , Boldú L , Molina Á , Alférez S , Rodellar J . A new convolutional neural network predictive model for the automatic recognition of hypogranulated neutrophils in myelodysplastic syndromes . Comput Biol Med 2021 ; 134 : 104479 . OpenUrl PubMed Saraswat M , Arya KV . Automated microscopic image analysis for leukocytes identification: a survey . Micron 2014 ; 65 : 20 – 33 . OpenUrl PubMed Wu Y-Y , Huang T-C , Ye R-H , Fang W-H , Lai S-W , Chang P-Y et al. A Hematologist-Level Deep Learning Algorithm (BMSNet) for Assessing the Morphologies of Single Nuclear Balls in Bone Marrow Smears: Algorithm Development . JMIR Med Inform 2020 ; 8 : e15963 . OpenUrl Rezatofighi SH , Soltanian-Zadeh H . Automatic recognition of five types of white blood cells in peripheral blood . Comput Med Imaging Graph 2011 ; 35 : 333 – 343 . OpenUrl CrossRef PubMed Mori J , Kaji S , Kawai H , Kida S , Tsubokura M , Fukatsu M et al. Assessment of dysplasia in bone marrow smear with convolutional neural network . Sci Rep 2020 ; 10 : 14734 . OpenUrl PubMed ↵ Fu X , Fu M , Li Q , Peng X , Lu J , Fang F et al. Morphogo: An Automatic Bone Marrow Cell Classification System on Digital Images Analyzed by Artificial Intelligence . Acta Cytol 2020 ; 64 : 588 – 596 . OpenUrl CrossRef PubMed ↵ Eckardt J-N , Middeke JM , Riechert S , Schmittmann T , Sulaiman AS , Kramer M et al. Deep learning detects acute myeloid leukemia and predicts NPM1 mutation status from bone marrow smears . Leukemia 2022 ; 36 : 111 – 118 . OpenUrl CrossRef PubMed ↵ Eckardt J-N , Schmittmann T , Riechert S , Kramer M , Sulaiman AS , Sockel K et al. Deep learning identifies Acute Promyelocytic Leukemia in bone marrow smears . BMC Cancer 2022 ; 22 : 201 . OpenUrl PubMed ↵ Röllig C , Thiede C , Gramatzki M , Aulitzky W , Bodenstein H , Bornhäuser M et al. A novel prognostic model in elderly patients with acute myeloid leukemia: results of 909 patients entered into the prospective AML96 trial . Blood 2010 ; 116 : 971 – 978 . OpenUrl Abstract / FREE Full Text ↵ Schaich M , Parmentier S , Kramer M , Illmer T , Stölzel F , Röllig C et al. High-dose cytarabine consolidation with or without additional amsacrine and mitoxantrone in acute myeloid leukemia: results of the prospective randomized AML2003 trial . J Clin Oncol 2013 ; 31 : 2094 – 2102 . OpenUrl Abstract / FREE Full Text ↵ Röllig C . Intermediate-dose cytarabine plus mitoxantrone versus standard-dose cytarabine plus daunorubicin for acute myeloid leukemia in elderly patients . Ann Oncol 2018 ; 29 : 973 – 978 . OpenUrl PubMed ↵ Röllig C , Serve H , Hüttmann A , Noppeney R , Müller-Tidow C , Krug U et al. Addition of sorafenib versus placebo to standard therapy in patients aged 60 years or younger with newly diagnosed acute myeloid leukaemia (SORAML): a multicentre, phase 2, randomised controlled trial . Lancet Oncol 2015 ; 16 : 1691 – 1699 . OpenUrl CrossRef PubMed ↵ Arber DA , Orazi A , Hasserjian RP , Borowitz MJ , Calvo KR , Kvasnicka H-M et al. International Consensus Classification of Myeloid Neoplasms and Acute Leukemias: integrating morphologic, clinical, and genomic data . Blood 2022 ; 140 : 1200 – 1228 . OpenUrl CrossRef PubMed ↵ Parmentier S , Kramer M , Weller S , Schuler U , Ordemann R , Rall G et al. Reevaluation of reference values for bone marrow differential counts in 236 healthy bone marrow donors . Ann Hematol 2020 ; 99 : 2723 – 2729 . OpenUrl PubMed ↵ World Medical Association . World Medical Association Declaration of Helsinki: ethical principles for medical research involving human subjects . JAMA 2013 ; 310 : 2191 – 2194 . OpenUrl CrossRef PubMed Web of Science ↵ Swerdlow , SH , Campo , E , Harris , NL , Jaffe , ES , Pileri , SA , Stein , H et al. WHO Classification of Tumours of Haematopoietic and Lymphoid Tissues . https://publications.iarc.fr/Book-And-Report-Series/Who-Classification-Of-Tumours/WHO-Classification-Of-Tumours-Of-Haematopoietic-And-Lymphoid-Tissues-2017 (accessed 19 Nov2024). ↵ He K , Zhang X , Ren S , Sun J. Deep Residual Learning for Image Recognition . In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE : Las Vegas, NV, USA , 2016 , pp 770 – 778 . ↵ Xie S , Girshick R , Dollár P , Tu Z , He K. Aggregated Residual Transformations for Deep Neural Networks . In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . 2017 , pp 5987 – 5995 . ↵ Zagoruyko S , Komodakis N. Wide Residual Networks . 2017 . doi: 10.48550/arXiv.1605.07146 . OpenUrl CrossRef ↵ Huang G , Liu Z , Maaten L van der , Weinberger KQ. Densely Connected Convolutional Networks . 2018 . doi: 10.48550/arXiv.1608.06993 . OpenUrl CrossRef ↵ Zhang X , Zhou X , Lin M , Sun J. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices . 2017 . doi: 10.48550/arXiv.1707.01083 . OpenUrl CrossRef ↵ Iandola FN , Han S , Moskewicz MW , Ashraf K , Dally WJ , Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size . 2016 . doi: 10.48550/arXiv.1602.07360 . OpenUrl CrossRef ↵ Deng J . ImageNet: A large-scale hierarchical image database . In: 2009 IEEE Conference on Computer Vision and Pattern Recognition 248–255 . 2009 doi: 10.1109/CVPR.2009.5206848 . OpenUrl CrossRef ↵ Akiba T , Sano S , Yanase T , Ohta T , Koyama M. Optuna: A Next-generation Hyperparameter Optimization Framework . In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . Association for Computing Machinery : New York, NY, USA , 2019 , pp 2623 – 2631 . ↵ Bain BJ . Diagnosis from the blood smear . N Engl J Med 2005 ; 353 : 498 – 507 . OpenUrl CrossRef PubMed Web of Science ↵ Lee N , Jeong S , Park M-J , Song W . Deep learning application of the discrimination of bone marrow aspiration cells in patients with myelodysplastic syndromes . Sci Rep 2022 ; 12 : 18677 . OpenUrl PubMed ↵ Kimura K , Tabe Y , Ai T , Takehara I , Fukuda H , Takahashi H et al. A novel automated image analysis system using deep convolutional neural networks can assist to differentiate MDS and AA . Sci Rep 2019 ; 9 : 13385 . OpenUrl PubMed ↵ Castelvecchi D . Can we open the black box of AI? Nature 2016 ; 538 : 20 – 23 . OpenUrl CrossRef PubMed ↵ Joshi G , Walambe R , Kotecha K. A Review on Explainability in Multimodal Deep Neural Nets . 2021 . doi: 10.48550/arXiv.2105.07878 . OpenUrl CrossRef ↵ Malcovati L , Karimi M , Papaemmanuil E , Ambaglio I , Jädersten M , Jansson M et al. SF3B1 mutation identifies a distinct subset of myelodysplastic syndrome with ring sideroblasts . Blood 2015 ; 126 : 233 – 241 . OpenUrl Abstract / FREE Full Text ↵ Papaemmanuil E , Cazzola M , Boultwood J , Malcovati L , Vyas P , Bowen D et al. Somatic SF3B1 mutation in myelodysplasia with ring sideroblasts . N Engl J Med 2011 ; 365 : 1384 – 1395 . OpenUrl CrossRef PubMed Web of Science ↵ Brück OE , Lallukka-Brück SE , Hohtari HR , Ianevski A , Ebeling FT , Kovanen PE et al. Machine Learning of Bone Marrow Histopathology Identifies Genetic and Clinical Determinants in Patients with MDS . Blood Cancer Discovery 2021 ; 2 : 238 – 249 . OpenUrl Abstract / FREE Full Text ↵ Nagata Y , Zhao R , Awada H , Kerr CM , Mirzaev I , Kongkiatkamon S et al. Machine learning demonstrates that somatic mutations imprint invariant morphologic features in myelodysplastic syndromes . Blood 2020 ; 136 : 2249 – 2262 . OpenUrl PubMed ↵ Steensma DP . Dysplasia has A differential diagnosis: distinguishing genuine myelodysplastic syndromes (MDS) from mimics, imitators, copycats and impostors . Curr Hematol Malig Rep 2012 ; 7 : 310 – 320 . OpenUrl CrossRef PubMed ↵ Iolascon A , Heimpel H , Wahlin A , Tamary H . Congenital dyserythropoietic anemias: molecular insights and diagnostic approach . Blood 2013 ; 122 : 2162 – 2166 . OpenUrl Abstract / FREE Full Text ↵ Huff JD , Keung Y-K , Thakuri M , Beaty MW , Hurd DD , Owen J et al. Copper deficiency causes reversible myelodysplasia . Am J Hematol 2007 ; 82 : 625 – 630 . OpenUrl CrossRef PubMed Web of Science ↵ Batata M , Spray GH , Bolton FG , Higgins G , Wollner L . Blood and bone marrow changes in elderly patients, with special reference to folic acid, vitamin B12, iron, and ascorbic acid . Br Med J 1967 ; 2 : 667 – 669 . OpenUrl FREE Full Text ↵ Sheikha A . Dyserythropoiesis in 105 patients with visceral leishmaniasis . Lab Hematol 2004 ; 10 : 206 – 211 . OpenUrl PubMed ↵ Michot F , Gut J . Alcohol-induced bone marrow damage. A bone marrow study in alcohol-dependent individuals . Acta Haematol 1987 ; 78 : 252 – 257 . OpenUrl CrossRef PubMed Web of Science ↵ Dusse LMS , Moreira AMB , Vieira LM , Rios DRA , Silva RMM e , Carvalho M das G . Acquired Pelger-Huët: what does it really mean? Clin Chim Acta 2010 ; 411 : 1587 – 1590 . OpenUrl CrossRef PubMed Web of Science View the discussion thread. Back to top Previous Next Posted January 28, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia Jan-Niklas Eckardt , Ishan Srivastava , Freya Schulze , Susann Winter , Tim Schmittmann , Sebastian Riechert , Martin M. K. Schneider , Lukas Reichel , Miriam Eva Helena Gediga , Katja Sockel , Anas Shekh Sulaiman , Christoph Röllig , Frank Kroschinsky , Anne-Marie Asemissen , Christian Pohlkamp , Torsten Haferlach , Martin Bornhäuser , Karsten Wendt , Jan Moritz Middeke medRxiv 2025.01.27.25321165; doi: https://doi.org/10.1101/2025.01.27.25321165 Share This Article: Copy Citation Tools Image-based Explainable Artificial Intelligence Accurately Identifies Myelodysplastic Neoplasms Beyond Conventional Signs of Dysplasia Jan-Niklas Eckardt , Ishan Srivastava , Freya Schulze , Susann Winter , Tim Schmittmann , Sebastian Riechert , Martin M. K. Schneider , Lukas Reichel , Miriam Eva Helena Gediga , Katja Sockel , Anas Shekh Sulaiman , Christoph Röllig , Frank Kroschinsky , Anne-Marie Asemissen , Christian Pohlkamp , Torsten Haferlach , Martin Bornhäuser , Karsten Wendt , Jan Moritz Middeke medRxiv 2025.01.27.25321165; doi: https://doi.org/10.1101/2025.01.27.25321165 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Hematology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (297) Cardiovascular Medicine (4421) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (606) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15212) Forensic Medicine (30) Gastroenterology (1121) Genetic and Genomic Medicine (6581) Geriatric Medicine (667) Health Economics (996) Health Informatics (4520) Health Policy (1366) Health Systems and Quality Improvement (1611) Hematology (539) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15906) Intensive Care and Critical Care Medicine (1103) Medical Education (620) Medical Ethics (144) Nephrology (667) Neurology (6580) Nursing (345) Nutrition (998) Obstetrics and Gynecology (1141) Occupational and Environmental Health (956) Oncology (3324) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1689) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5432) Public and Global Health (9212) Radiology and Imaging (2193) Rehabilitation Medicine and Physical Therapy (1368) Respiratory Medicine (1194) Rheumatology (593) Sexual and Reproductive Health (709) Sports Medicine (529) Surgery (709) Toxicology (99) Transplantation (288) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff47acd2cd04807',t:'MTc3OTM3NTkzMA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();