Indonesian J our nal of Electrical Engineering and Computer Science V ol. 38, No. 2, May 2025, pp. 950 ∼ 959 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v38.i2.pp950-959 ❒ 950 Acute lymphoblastic leuk emia diagnosis and subtype segmentation in blood smears using CNN and U-Net Hamim Reza 1 , Nazrul Islam T ar eq 1 , M M F azle Rab bi 1 , Sharia Arn T anim 2 , Rifat Al Mamun Rudr o 2 , Kamruddin Nur 2 1 Department of Computer Science and Engineering, Bangladesh Uni v ersity of Business and T echnology , Dhaka, Bangladesh 2 Department of Computer Science, American International Uni v ersity-Bangladesh, Dhaka, Bangladesh Article Inf o Article history: Recei v ed Jun 10, 2024 Re vised Oct 28, 2024 Accepted No v 11, 2024 K eyw ords: Acute lymphoblastic leuk emia CNN Se gmentation Blood Smears Hematogone ABSTRA CT Acute lymphoblastic leukaemia (ALL) is a se v ere disease requiring in v asi v e, e xpensi v e, and time-consuming di agnostic tests for deniti v e diagnosis. Initial diagnosis using blood smear pictures (BSP) is crucial b ut challenging due to the similar indications and symptoms of ALL, often leading to misdiagnoses. This study presents a custom approach using Con v oluti onal Neural Netw orks (CNNs) to detect all cases and cate gorize subtypes. Utilizing publicly a v ailable databases, the study includes 3562 blood smear images from 89 patients. The in- no v ati v e combination of U-Net for se gmentation and v arious CNN architectures (U-Net, MobileNetV2, InceptionV3, ResNet50, N ASNet) for feature e xtraction, with DenseNet201 being the most ef fecti v e, forms the core of this method. The U-Net model achie v ed a se gmentation ac curac y of 98% by recognizing patterns within blood smear images. F ollo wing se gmentation, CNN architectures e x- tracted high-le v el features, with DenseNet201 pro ving the most ef fecti v e in di- agnostic and classication tasks. Our proposed custom CNN model achie v ed a test accurac y of 98%, wi th a training accurac y of 99.31% and v alidation ac- curac y of 97.09%. This approach enables an accurate distinction between ALL and non-pathologic cases. This is an open access article under the CC BY -SA license . Corresponding A uthor: Kamruddin Nur Department of Computer Science, American International Uni v ersity-Bangladesh Dhaka, Bangladesh Email: kamruddin@aiub .edu 1. INTR ODUCTION Acute lymphoblastic leuk emia (ALL) is a highly common type of cancer that requires careful and sometimes in v asi v e diagnostic methods to identify it accurately . Precise identication of ALL especially dur - ing its initial phases, is crucial for prompt interv ention and ef cient treatment. Peripheral blood smear (PBS) [1] images are highly important diagnostic tools that pro vide v aluable information about cellular abnormalities that indicate the presence of leukaemia. The manual reading of PBS [2] images to e xplore decided disease issues is af fected by the v ast problems related to the risk of wrong diagnosis that may result from the lo w and ambiguous features of the patient’ s signs. Bad interpretations could cause patients to appear too often, leading to misdiagnosis and less ef fecti v e treatments, xing the w ors t-case situations and increasing the b urden on the healthcare system. Here with proposed is an approach to creating a cutting-edge tool which will enable the classication and precis e diagnosis of all breakpoints and ALL subtypes e xploiting cutting-edge deep learning approaches. This project aims to apply ALL detecti on to PBS pictures more automatically and to di vide the J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 951 ALL [3] cases into benign and INFMUN. The whole project goal is to de v elop a thorough and easily man- ageable dataset that contains pictures from people diagnosed with ALL [4], acute lymphoid leukaemia. This dataset presents dif ferent types of scenarios including those that are considered harmless from the population haematogone syndrome as wel l as the conrmed occurrence of the ALL subtypes [5], which i n turn gi v es a chance for an ef fecti v e assessment of the training model and e v aluation. In addition to that, the grouping of PBS [6] images using dif ferent hue threshold techniques in the HSV colour space, percei v ed as the preliminary step for the meticulous feature e xtraction, lays the foundat ion for a high le v el of accurac y in the prediction. This research mak es a comparati v e study of CNN architectures that are cataloged as the best ones, that is, U-Net, MobileNetV2, InceptionV3, ResNet50 and V iT and N ASNet. The results demonstrate the best of Dens eNet201 performance (diagnosis and classication tasks) that is e vidently superior to the others. The k e y contrib utions of this study are: − Introduce a no v el approach for ALL diagnoses by con v olutional neural netw orks (CNNs) for precise classication. − Utilization of PBS images for reducing the risk of misdiagnosis associated with manual interpretation. − The proposed model signicantly impro v es clinical specicity , enabling a reliable diagnosis of ALL. The subsequent sections of the paper are structured in the follo wing manner: Section 2. discusses t he w ork of my predecessors. In Section 3. e xplanation the methodology step by step. The model outcome and the primiti v e actions that should be tak en as an outcome are discussed in Section 4. Section 5. concl ude the paper . 2. RELA TION W ORK ALL classication using CNN and transfer learning has sho wn promising results in impro ving ef - cienc y and accurac y in identifying leuk emia cells. Khuzaie et al. [7] discusses using a V GG19-based CNN model for detecting ALL cells. The paper focuses on de v eloping an ef cient V GG19-based model for detect- ing ALL. Deep learning t echniques can streamline the identication of leukaemia cells and impro v e patient outcomes. Das et al. [8] proposes a model for classifying and detecting ALL using transfer learning and an orthogonal SoftMax layer (OSL)-based classication. Demonstrates superior performance on ALLIDB1, AL- LIDB2, and CNMC2019 datasets.The paper proposes a model for detecting and classifying acute leuk emia. The model combines ResNet18 with an orthogonal SoftMax layer for impro v ed performance. Ahammed et al. [9] proposes an ef cient transfer -learning-based CNN model using Inception-V3 architecture to classify ALL from microscopic images. Also, Hau et al. [10] proposes a h ybrid transfer learning eXtreme gradient boosting (HTL-XGB) algorithm for the classication and detection of ALL using CNNs and transfer learning. Object detection methodology using image processing techniques with HTL-XGB architecture. Gautam et al. [11] introduced a classication method for WBCs that combines the Nai v e Bayes classier with morphological fea- tures. The characteristics the researchers emplo yed to train their system included area, perimeter , eccentricity , and circularity . The accurac y of the proposed method w as made up of 80.88 percent. The process of manually classifying acute lymphoblastic leuk emia is laborious and t ime-intensi v e. The proposed procedure emplo ying Mask R-CNN attains a classication accurac y of 83.72%. Se gmentation of instances utilizing mask R-CNN Method for enhancing contrast in an image dataset [12]. Recent in v estig ations into classifying malignancies ha v e relied hea vily on computer vision me thods [13]-[16]. The predominant algorithm ut ilized in this methodology comprises multiple e v aluations that fol- lo w image pre-processing, clustering, morphological ltering, se gmentation, feature e xtraction or selection, and classication [17]. These are rigid phases. Due to the diagnostic importance of blood cell classication, numerous algorithms for classifying blood cells ha v e been proposed by scientists. Sinha and Ramakrishnan V OLUME XX, 2018 classied cells with a 94.1% recognition rate using SVM in [18]. The researchers con- ducted the identical e xperiments using one hundred images. The researchers em plo yed the method with the smallest error rate to classify the se gmented cells using an adapti v e contour and automatic threshold. The resulting recogniti on rate w as 96%. The researchers put the KNN algorithm to use. Ne v ertheless, the KNN algorithm struggles to process unbalanced samples. Dif culties may arise when the sample capacity of one class is substantial, while that of other classes is relati v ely limited. Leuk emia causes premature death and other symptoms in children and adults. Computer -aided m eth- ods can help specialists diagnose this disease and pre v ent incorrect therap y prescriptions. CNNs [19] are Acute lymphoblastic leuk emia dia gnosis and subtype se gmentation in blood smear s using ... (Hamim Reza) Evaluation Warning : The document was created with Spire.PDF for Python.
952 ❒ ISSN: 2502-4752 increasingly used to classify and diagnose medical images. Ho we v er , CNN training in v olv es man y images. W e emplo y transfer learning to e xtract picture features for classication to solv e this challenge. Leuk emia is a deadly white blood cell i llness that af fects blood and bone marro w . Deep con v ol utional neural netw ork w as used to detect acute lymphoblastic leuk emia and classify its subtypes into four classes: L1, L2, L3, and Nor - mal, which were disre g arded in prior studies. Instead of training from scratch, we used pre-trained Ale xNet [20] ne-tuned on our data set. Ne w layers classify incoming photos into four classications, replacing the pretrained netw ork’ s last le v els. Ov ertraining w as reduced via data augmentation. 3. METHOD The follo wing section pro vides a comprehensi v e o v ervie w of the e xperimental frame w ork. Ini tially , we ha v e to choose the dataset and then use the data pre-processing techniques to t the data for the model. W e conducted e xperiments by using publicly acces sible data. The Zeiss microscope at 100x magnication w as used to capture blood smear images, which were sa v ed in JPG format. T o mak e it adaptable for the deep learning model, we ha v e standardized images into 224x224 pix els through preprocessing techniques. This preprocessing included applying rotation, contrast adjustment, and se gmentation in the HSV colour space. After the pre- processing data, we implemented the deep learning models and e v aluated the result on the e v aluation metrics. This section also pro vides the dataset description, a concise analysis of the deep learning models, and an e v aluation of the proposed system’ s performance. The fundamental architecture of our research is illustrated in Figure 1. Figure 1. Methodology w orko w for acute lymphoblastic leuk emia detection using augmented data, a CNN model, and performance e v aluation metrics 3.1. Dataset F or disease detection of ALL, we are using a dataset [21] of 3,256 images that ha v e been tak en from 89 ALL patients who did P BS [22]-[24] e xamination at T aleqani Hospital in T ehran, Iran. The images were di vided into t w o classes: harmful with a benign w ay tend to ha v e the capabilit y of teari ng and destr o y i ng important molecules from a cell and self-protection of the body ag ainst cancerous action. The malignant class contained three sub-types of malignant lymphoblasts: T able 1 pre sents the early Pre-B, Pre-B, and Pro-B ghting ALL [25] of this joint ef fort of modern medicine in T able 1. T able 1. Dataset distrib ution for diagnosing ALL Class name T otal image T rain V alidation T est Benign 504 323 80 101 Malignant-early 985 631 157 197 Malignant-pre (Pre-B) 963 616 154 193 Malignant-pro (Pro-B) 804 516 128 160 W e ha v e tak en all images with a Zeiss camera with 100x magnication and sa v ed t hem as JPG les. A specialist deniti v ely determined the cell types and subtypes using the o w c ytometry tool. In addition, se gmented images were pro vided after applying colour thresholding-based se gmentation in the HSV colour space, sho wn in T able 2. Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 950–959 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 953 T able 2. Acute lymphoblastic leuk emia dataset details Feature Details Dataset origin Bone marro w laboratory , T aleqani Hospital, T ehran, Iran T otal images 3256 peripheral blood smear (PBS) images Number of patients 89 suspected of ALL Preparation Prepared and stained by skilled laboratory staf f Image format JPG les Imaging equipment Zeiss camera, microscope at 100x magnication Diagnosis conrmation Specialist using o w c ytometry Se gmentation technique Color thresholding in HSV color space; se gmented images pro vided ALL subtypes Hematogones, early Pre-B, Pre-B, Pro-B ALL 3.2. Dataset pr e-pr ocessing After completi o n of datas et selection and import for the implementation, we maintain the origi nal ratio, and achie ving consistent data distrib ution through normalization is crucial during the pre-processing stage of image data. T o ac hie v e this, we establish a x ed tar get size parameter of 224x224 pix els, ensuring that all images loaded into the deep-learning model are resized to this size. This is essential since deep-learning models typically require data of a specic size. By standardizing the image shape, we enable the model to process the data ef ciently and with precision. The pre-processing techniques emplo yed in our res earch includes rotation, contrast, ipping, cropping, cutout and brightness pre-processing. 3.3. Deep lear ning models 3.3.1. Customized CNN ar chitectur e The deep learning model designed for image recognition tasks has a customized CNN architec ture. The model includes an input layer specically for RGB images of size 224x224 in Figure 2. It also has con- v olutional layers for feature e xtraction, pooling layers for dimensionality reduction, dropout layers to pre v ent o v ertting, and dense layers for classication. The ReLU acti v ation, batch normalization, and dropout tech- niques are emplo yed in the model’ s architecture to impro v e its performance and generalization. − Input layer: (224, 224, 3). The model has a model card that e xpects an image of resoluti on 224x224 with 3 channels (RGB). − Con v olutional layers: the model consists of a con v olutional layer of four layers arranged in the sequence and with the ReLU acti v ation function which introduces non-linearity . There are 200 lters in the rst con v olutional layer , 150 lters in the second con v olutional layer , and so on do wn to 50 lters in the fourth con v olutional layer with each layer 3x3 k ernel sizes. The idea of this structure allo ws the netw ork to e xtract characteristics of the image at dif ferent de grees of abstraction. − Pooling layers: the rst and second layers are max-pooling layers with a pool size of (4,4). − Dropout layers: applied twice with a rate of 0.8, after the rst and third con v olutional layers to pre v ent o v ertting. − Flatten layer: this layer transforms the 2D arrays from the pre vious layers into a 1D array , preparing it for the fully connected layers. − Fully connected (Dense) layers: after the con v olutional layers, the architecture includes a dense layer with 256 units and ReLU acti v ation, follo wed by batch normalization for stabilized acti v ation and a dropout rate of 0.8 to reduce o v ertting. Figure 2. A customized CNN model with con v olutional layers, pooling, dropout, attening, and fully connected layers, culminating in a 4-unit output for classication Acute lymphoblastic leuk emia dia gnosis and subtype se gmentation in blood smear s using ... (Hamim Reza) Evaluation Warning : The document was created with Spire.PDF for Python.
954 ❒ ISSN: 2502-4752 3.3.2. U-Net ar chitectur e Contracting path (encoder): the architecture in v olv es a contracting functional block (encoder) and a transformational functional block (decoder). The contracting passage looks lik e a simple CNN architectural design that comprises se v eral con v olutional and pooling layers. Each contracting path block, lik e all the others, generally has tw o 3x3 con v olutions, follo wed by a rectied linear unit (ReLU ) acti v ation function and max- pooling with W indo ws sub-sampling size 2x2. This helps capture conte xt and reduce the spatial dimensions sho wn in Figure 3. Figure 3. V isualization of U-net model architecture Expansi v e path (decoder): the e xpansi v e path-up samples feature maps to the original input size, increases resolution, and reco v ers spatial information lost during do wnsampling. It uses up-con v olutional layers (transposed con v olutions or decon v olutions) to boost spatial resolution. Concatenation of contracting and e xpansi v e path feature maps pro vides e xtensi v e localization information. Final layer: the last layer is the most important layer as follo ws: it consists of a 1x1 con v olutional layer and a soft-max acti v ation function with an output of the se gmentation mask composed of pix el-wise classication probabilities from each class. At the end of the netw ork channels present a number equals a number of se gmentation classes. 3.4. Ev aluation metrics The models were e v aluated based on the accurac y in (1), precision in (2), recall in (3), and F1-score in (4). Accur acy = T P + T N T P + T N + F P + F N (1) P r ecision = T P T P + F P (2) R ecal l = T P T P + F N (3) F 1 scor e = 2 × precision × recall precision + recall (4) In (5), X represents the pix els in the predict ed se gmentation, and Y denes the pix els in the ground truth se gmentation. D iceC oef f . = 2 × | X ∩ Y | | X | + | Y | (5) Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 950–959 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 955 4. RESUL T AND DISCUSSION 4.1. Experiment r esults The results obt ained by means of the proposed CNN settings are ob viousl y higher than an y other algorithms in terms of accurac y , precision, recall, and F1 score. When compared to the other approaches, the algorithm consistently outperforms them across all e v aluation metrics presented in Figure 4. Figure 4. T raining and v alidation metrics of customized CNN o v er epochs 4.1.1. U-Net model segmentation r esult Our dataset sho wed that the U-Net model is ef fecti v e in precise se gmentation. It can accurately i d e n- tify patterns, as e videnced by its consistently impro ving binary accurac y . The model is capable of generalizing ef fecti v ely to ne w data, which is crucial for real-w orld medical applications, as demonstrated by its v alida- tion accurac y of 0.8890 in T able 3. The test results were slightly conserv ati v e b ut still solid, with accuracies ranging from 0.8206 to 0.8446. The U-Net model’ s dice coef cient impro v ed signicantly , peaking at 0.5215, demonstrating it s precision in se gmentation in Figure 5 tasks for accurate medical diagnosis a n d interv ention planning. T able 3. Model training and test metrics o v er epochs Epoch T raining metrics T est metrics Loss Dice coef Binar y accurac y V al binary accurac y T est accurac y 1 0.7000 0.3000 0.6001 0.5914 0.5800 2 0.5691 0.4304 0.8502 0.7835 0.7700 3 0.5270 0.4730 0.8646 0.8251 0.8100 4 0.5117 0.4880 0.8681 0.8469 0.8300 5 0.0499 0.5004 0.8684 0.8638 0.8206 6 0.4929 0.5068 0.8722 0.8544 0.8117 7 0.4936 0.5062 0.8744 0.8866 0.8423 8 0.4784 0.5215 0.8750 0.8885 0.8441 9 0.4839 0.5161 0.8797 0.8890 0.8446 Figure 5. U-Net se gmentation for input image, true mask, and predicted mask Acute lymphoblastic leuk emia dia gnosis and subtype se gmentation in blood smear s using ... (Hamim Reza) Evaluation Warning : The document was created with Spire.PDF for Python.
956 ❒ ISSN: 2502-4752 4.1.2. Deep lear ning model classication r esults Precision, recall, and F-score metrics are three approaches to measuring the accurac y of the class ier . Precision estimates the classier’ s ability to nd correct positi v e cases from the real ones. This is calculated by comparing the number of genuine positi v es with the e xpected number of positi v es. The ability of the classier is judged by ho w man y of true-positi v e cases of the actual positi v e class are identied correctly by it and it is called as recall. The estimator of recall is the v alue obtained after di viding the number of the true positi v es by the total number of the actual positi v e cases. Quantifying true positi v e rate for capturing all positi v e samples is mark ed as a performance metric of the classier . The F 1-score, computed as the harmonic mean of precision and recall, which is well-balanced between the performance of each model, is the measure used. 4.2. Results analysis T able 4 sho ws a comparati v e model-by-model analysis which is focused on each model performance across dif ferent metrics, e.g. training accurac y , v alidation accurac y , precision, recall and F1 score. The tailored CNN and ResNet50 are no w the most promising, ha ving attained the highest classication scores ef fecti v e in real-life applications. In addition to MobileNet V2’ s ef fecti v eness, its computational ef cienc y should be mentioned e xplicitly . Look at InceptionV3 and NasNet, which ha v e signicantly lo wer performance and may need renement to achie v e ef cienc y . T able 4. Comparison of model performances Model name T raining accurac y V alidation accurac y Precision Recall F1-Score T est accurac y MobileNetV2 97.87 % 98.90 % 98.50 % 97.00 % 98.00 % 97.00 % InceptionV3 77.69 % 82.52 % 82.00 % 77.00 % 76.00 % 77.00 % ResNet50 98.04 % 98.96 % 98.60 % 98.00 % 96.00 % 97.00 % NasNet 83.76 % 84.82 % 84.30 % 82.00 % 83.00 % 81.00 % Customized CNN 99.31 % 97.09 % 96.80 % 97.00 % 99.00 % 98.00 % In Figure 6 the customized CNN achie v ed high training and v alidation accurac y rates, which means it remained great on the test and carried out the task perfectly . On this it established a harmon y in terms of nding all delightful e xamples with 97% precision, 99% recall, and 98% F1 score. MobileNetV2 had lo wer accurac y in v alidation and training than a customized CNN, b ut it still managed to achie v e the desirable accurac y v alues of 97% precision, recall, and F1 score. Such a model is well-trainable and can be a good option for tasks that require high comple xity of the model, b ut at the same time, high accurac y of the performance is needed. Figure 6. T raining, v alidation, and test accurac y of customized CNN o v er epochs InceptionV3, which w as characterized by lo wer training and v alidation accurac y implications, has a major a w in learning. Its accurac y v alues were also lo wer , and thus, the recall w as lo wer , resulting in the F1 quantity being 0.77. Using well-tuned h yperparameters, more data or enlar ging the training sequence may also help. ResNet50 got good training and v alidation accuracies which is an indication of its good performance, sho wing precision of almost 98% b ut only 96% recall. This sho wed a 97% F-1 score, which also i ndicated a slight tendenc y t o w ards precision. It can nd positi v e cases quick er b ut might not be able to do so completely (true positi v es par t) with the customize CNN. NasNet, achie v ed the results of a moderate le v el because of its lo w learning and v oting accuracies. Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 950–959 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 957 5. CONCLUSION This article pro vides a comprehensi v e e xamination of v arious deep learning architectures for clas - sication tas ks, emphasizing the importance of model selection and optimization to achie v e high accurac y and generalizability . Our analysis re v ealed that the CNN achie v ed outs tanding performance with almost full accurac y by the end of training, reaching 99% accurac y . This lo w error rate suggests its suitability for de- plo yment in real w orld scenarios. The CNN-C model achie v ed a training accurac y of 31% and a v alidation accurac y of 99.74%, outperforming other models in solving the classication task. The MobileNetV2 model also demonstrated rob ust performance, with a training a ccurac y of 97.87% and a v alidation accurac y of 98%, and a precision of 90%, making it ef fecti v e and accurate, especially in resource-constrained en vironments. In contrast, the Ince ptionV3 and NasNet models sho wed more modest results, with InceptionV3 achie ving a training accurac y of 77% and NasNet achie ving 75.49% accurac y . The v alidation accurac y for NasNet w as 82%, higher than man y recent studies. Models lik e V GG, ResNet, and Inception Net demonstrated a training accurac y of 52%, while NasNet sho wed an accurac y of 83% in training. The o v erall test dataset accurac y w as 76%, and the v alidation dataset accurac y w as 84%, indicating the potential for further ne-tuning and opti- mization. Notably , the ResNet50 model achie v ed a training accurac y of 98%, highlighting its ef fecti v eness. F or image classication tasks, the customized CNN closely matched the performance of other deep learning models, demonstrating the competiti v eness of deep learning in this domain. A CKNO WLEDGMENT The authors w ould lik e to ackno wledge the UCH Research Group, American International Uni v ersi ty- Bangladesh, and the Bangladesh Uni v ersity of Business and T echnology for supporting this collaborati v e re- search. REFERENCES [1] S. Perv een, A. Alourani, M. Shahbaz, M. U. Ashraf, and I. Hamid, “ A frame w ork for early detect ion of acute lymphoblastic leuk emia and its subtypes from peripheral blood smear images using deep ensemble learning technique, ” IEEE Access , v ol. 12, pp. 29 252–29 268, 2024. [2] F . Akalin and N. Y umus ¸ ak, “Early detection of all disease using yolo v4 algorithm on peripheral blood smear images, ” in 2022 Inno vations in Intellig ent Systems and Applications Confer ence (ASYU) , 2022, pp. 1–5, doi: 10.1109/ASYU56188.2022.9925427. [3] N. Gopig ari and T . Singh, “Comparison of se gmentation techniques for acute lymphoblastic leuk emia in leuk emia cancer , ” in 2022 13th International Confer ence on Computing Communication and Netw orking T ec hnolo gies (ICCCNT) , 2022, pp. 1–5, doi: 10.1109/ICCCNT54827.2022.9984527. [4] Z. Boreiri, A. N. Azad, and A. Ghodousian, “ A con v olutional neuro-fuzzy netw ork using fuzzy image se gmentation for acute leuk emia classication, ” in 2022 27th International Computer Confer ence , Computer Society of Ir an (CSICC) , 2022, pp. 1–7, doi: 10.1109/CSICC55295.2022.9780525. [5] T . Mustaqim, C. F atichah, and N. Suciati, “Deep learning for the detection of acute lymphoblastic leuk emia subtypes on microscopic images: A systematic literature re vie w , ” IEEE Access , v ol. 11, pp. 16 108–16 127, 2023, doi: 10.1109/A CCESS.2023.3245128. [6] L. D. L and G. V , “Leuk emia diagnosis made easy: A machine learning solution using blood smear images, ” in 2023 3r d International Confer ence on Mobile Networks and W ir eless Communications (ICMN WC) , 2023, pp. 1–4, doi: 10.1109/ICM- NWC60182.2023.10435943. [7] M. Y . Al-khuzaie, S. A. Zearah, and N. J. Mohammed, “De v eloping an ef cient vgg19-based model and transfer learning for detecting acute lymphoblast ic leuk emia (all), ” in 2023 5th International Congr ess on Human-Computer Inter action, Optimization and Robotic Applications (HORA) , 2023, pp. 1–5, doi: 10.1109/HORA58378.2023.10156679. [8] P . Das, B. Sahoo, and S. Meher , “ An ef cient detection and classication of acute leuk emia using transfer learning and orthog- onal softmax layer -based model, ” IEEE/A CM T r ansactions on Computational Biolo gy and Bioinformatics , v ol. 20, no. 03, pp. 1817–1828, May 2023, doi: 10.1109/TCBB.2022.3218590. [9] P . Ahammed, M. F . F aruk, N. Raihan, and M. Mondal, “Inception v3 based transfer learning model for the prognosis of acute lymphoblastic leuk emia from microscopic images, ” in 2022 4th International Confer ence on Electrical, Computer & T elecommuni- cation Engineering (ICECTE) , 2022, pp. 1–4, doi: 10.1109/ICECTE57896.2022.10114522. [10] A. J. Hau, N. Hameed, A. W alk er , and M. M. Hasan, “ A h ybrid transfer learning and se gmentation approach for the detection of acute lymphoblastic leuk emia, ” in Pr oceedings of T r ends in Electr onics and Health Informatics , M. Mahmud, C. Mendoza-Barrera, M. S. Kaiser , A. Bandyopadh yay , K. Ray , and E. Lugo, Eds. Sing apore: Springer Nature Sing apore, 2023, pp. 175–189. [11] L. Xiao, Q. Y an, and S. Deng, “Scene classication with impro v ed ale xnet model, ” in 2017 12t h International Confer ence on Intellig ent Systems and Knowledg e Engineering (ISKE) , 2017, pp. 1–6, doi: 10.1109/ISKE.2017.8258820. [12] A. R. Re v anda, C. F atichah, and N. Suciati, “Classication of acute l ymphoblastic leuk emia on white blood cell microscop y images based on instance se gmentation using mask r -cnn, ” International J ournal of Intellig ent Engineering and Systems , v ol. 15, no. 5, p. 625, 2022, doi: 10.22266/ijies2022.1031.54. [13] R. Raina, N. K. Gondhi, Chaahat, D. Singh, M. Kaur , and H.-N. Lee, “ A systemati c re vie w on acute leuk emia detection using deep learning techniques, ” Ar c hives of Computational Methods in Engineering , v ol. 30, no. 1, pp. 251–270, 2023, doi: 10.1007/s11831- 022-09796-7. Acute lymphoblastic leuk emia dia gnosis and subtype se gmentation in blood smear s using ... (Hamim Reza) Evaluation Warning : The document was created with Spire.PDF for Python.
958 ❒ ISSN: 2502-4752 [14] P . K. Das and S. Meher , “T ransfer learning-based automatic detection of acute lymphoc ytic leuk emia, ” in 2021 National Confer ence on Communications (NCC) , 2021, pp. 1–6, doi: 10.1109/NCC52529.2021.9530010. [15] P . M. Sha, V . B idv e, H. Bhapkar , P . Dhotre, and V . B. P . Singh, “Leuk emia detection system using con v olutional neural net- w orks by means of microscopic pictures, ” Indonesian J ournal of Electrical Engineering and Computer Sciences , v ol. 31, no. 3, pp. 1616–1623, 2023, doi: 10.11591/ijeecs.v31.i3.pp1616-1623. [16] H. Abdulkarim, R. Sudirman, and M. A. Razak, “Normal and abnormal red blood cell recognition using image processing, ” Indone- sian J Elec Eng & Comp Sci , v ol. 14, no. 1, pp. 96–100, 2019, doi: 10.11591/ijeecs.v14.i1.pp96-100. [17] I. V incent, K.-R. Kw on, S.-H. Lee, and K.-S. Moon, “ Acute lymphoid leuk emia classication using tw o-step neural net- w ork classier , ” in 2015 21st K or ea-J apan J oint W orkshop on F r ontier s of Computer V ision (FCV) , 2015, pp. 1–4, doi: 10.1109/FCV .2015.7103739. [18] A. Rehman, N. Abbas, T . Saba, S. I. U. Rahman, Z. Mehmood, and H. K oli v and, “Classication of acute lymphoblastic leuk emia using deep learning, ” Micr oscopy Resear c h and T ec hnique , v ol. 81, no. 11, pp. 1310–1317, 2018, doi: 10.1002/jemt.23139. [19] L. H. V og ado, R. M. V eras, F . H. Araujo, R. R. Silv a, and K. R. Aires, “Leuk emia diagnosis in blood slides using transfer learning in cnns and svm for classication, ” Eng . Appl. Artif . Intell. , v ol. 72, pp. 415–422, 2018, doi: 10.1016/j.eng appai.2018.04.024. [20] S. Shaque and S. T ehsin, “ Acute lymphoblastic leuk emia detection and classication of its subtypes using pretrained deep con v o- lutional neural netw orks, ” T ec hnol. Cancer Res. T r eat. , v ol. 17, 2018, doi: 10.1177/1533033818802789. [21] M. Aria, M. Ghaderzadeh, D. Bashash, H. Abolghasemi, F . Asadi, and A. Hosseini, “ Acute lymphoblastic leuk emia (all) image dataset, ” in Ka g gle Dataset. Ka g gle , 2021, doi:10.34740/KA GGLE/DSV/2175623. [22] B. Ro y , M. F . F aruk, M. N. Islam, A. Y . Srizon, S. M. M. Hasan, M. A. Mamun, M. R. Hossain, and M. F . Hossain, “ A cutting-edge ensemble of vision transformer and resnet101v2 based transfer learning for the precise classication of leuk emia sub-types from peripheral blood smear images, ” in 2024 6th International Confer ence on Electrical Engineering and Information & Communication T ec hnolo gy (ICEEICT) , 2024, pp. 49–54, doi: 10.1109/ICEEICT62016.2024.10534388. [23] R. Satpathi, S. Das, M. Laha, and S. Ghosh, “ Acute L ymphoblastic Leuk emia Classication using Compact Channel Specic Multi- column CNNs, ” in 2023 8th International C onfer ence on Computer s and De vices for Communication (CODEC) , 2023, pp. 1–2, doi: 10.1109/CODEC60112.2023.10465891. [24] A. M. Basymeleh, B. E. Pramudya, and R. T . Santoso, “ Acute lymphoblastic leuk emia image classication performance with transfer learning using cnn architecture, ” in 2022 4th International Confer ence on Biomedical Engineering (IBIOMED) , 2022, pp. 30–35, doi: 10.1109/IBIOMED56408.2022.9988690. [25] S. Jain, P . V ishna w at, P . K. Shukla, and N. Khatri, “Detection of acute lymphoblastic leuk emia using collatenet, ” in 2023 3r d Inter - national Confer ence on T ec hnolo gical Advancements in Computational Sciences (ICT A CS) , 2023, pp. 1095–1100, doi: 10.1109/IC- T A CS59847.2023.10390121. BIOGRAPHIES OF A UTHORS Hamim Reza recei v ed his bachelor’ s de gree from Bangladesh Uni v ersity of Business and T echnology (B UBT). He is dedicated to continuous learning, al w ays approaching ne w chal- lenges with enthusiasm and passion. His research interests encompass machine learning, deep learning, pattern recognition, and image processing.His latest research paper represents his initial step into the w orld of academic research, sho wcasing his commitment to enhancing AI technolo- gies. He is currently w orking on research topics such as skin cancer detection and classication, and customized U-Net se gmentation on medical image datasets. He can be contacted at email: hamim.reza@cse.b ubt.edu.bd. Nazrul Islam T ar eq recently graduated from the Bangladesh Uni v ersity of Business and T echnology (B UBT). Dri v en by a k een interest in machine learning a nd deep learning, he focused his academic pursuits on the practical aspects of artici al intelligence. His latest research paper rep- resents his initial step into the w orld of academic research, sho wcasing his commitment to enhancing AI technologies. He is passionate about connecting theoretical AI concepts with their practical appli- cations to tackle real-w orld challenges. He can be contacted at email: nazrul.islam@cse.b ubt.edu.bd. M M F azle Rab bi has completed his MSc in Computer science from Uni v ersity of Bed- fordshire, UK and B.Sc de gree from Uni v ersity of W indsor , Canada. No w he is currently serving as assistant profess or in the department of CSE at Bangladesh Uni v ersity of Business and T echnology (B UBT). His research interests are machine learning, data science and IoT . He can be contacted at email: rabbi@b ubt.edu.bd. Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 950–959 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 959 Sharia Arn T anim is currently w orking as a Researcher at Learnify Research Lab (LRL) and as a Research Assistant (RA) at UCH Research Group. He graduated from the Department of Computer Science at the American International Uni v ersity-Bangladesh in December 2023. His research interests include decentralized learning methods, lar ge language models, computer vision, and pattern recognition. He can be contacted at email: shariaarn096@gmail.com. Rifat Al Mam un Rudr o is an alumnus and graduate researcher in the Department of Computer Science and Engineering at American International Uni v ersity-Bangladesh, deeply inter - ested in blockchain technology and articial intelligence. His research interests include deep learn- ing, AI, blockchain technology , rene w able ener gy . Additionally , he w orks as an instructor and is focused on research in deep learning and blockchain. Currently pursuing a master’ s in computer science specializing in data science, he is skilled in AI inte grat ion. He can be contacted at email: rif at.rudro@aiub .edu. Dr . Kamruddin Nur (Senior Member , IEEE) is currently serving as a full professor in the Department of Computer Science at American International Uni v ersity-Bangladesh (AIUB). He also serv ed as the chairman in the Department of Computer Science and Engineering at Stamford Uni v ersity Bangladesh and Bangladesh Uni v ersity of Busi ness and T echnology . Dr . Nur completed his PhD from UPF , Barcelona, Spain Masters from UIU, and Bachelor from V ictoria Uni v ersity of W ellington (VUW), Ne w Zealand. Dr . Nur authored ma n y prestigious journals and conferences in IEEE and A CM, serv ed as TPC members, and re vie wed articles in IEEE, A CM, Springer journals, and conferences. Currently , he is leading the ubiquitous, cloud Computing a nd HCI (UCH) research group at AIUB. His research area includes ubiquitous computing, computer vision, machine learning, and robotic automation. He can be contacted at email: kamruddin@aiub .edu. Acute lymphoblastic leuk emia dia gnosis and subtype se gmentation in blood smear s using ... (Hamim Reza) Evaluation Warning : The document was created with Spire.PDF for Python.