The accuracy of Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) in diagnosing different types of AcuteLymphocytic Leukemia based on Peripheral Blood Smear

Background and aims: ALL (Acute Lymphoblastic Leukemia) is the most prevalent type of leukemiain children and its screening and diagnosis process is often expensive, invasive, and needsto be done by experts. The initial screening process is commonly done by acquiring a PeripheralBlood Smear (PBS) from the suspected patient and examining the sample under a microscope.Further confirmation and diagnosis of the subtype are necessary for determining the right treatmentprotocol; which includes bone marrow aspiration and flow cytometry or specific moleculartests. Recently, new Artificial intelligence methods have provided more economical and availableways to diagnose different types of diseases. These methods are not only efficient in distinguishingbetween malignant and benign images, but also, they can classify subtypes of ALL. Thus, thisresearch intends to implement a CNN-LSTM (Convolutional Neural Network-Long Short TermMemory) model for classifying ALL subclasses.Method: In This Pipeline, we first divided our dataset into training (۹۰% of the whole dataset)and test (۱۰%) subsets, then preprocessed our data (Segmentation: RGB to LAB, Clustering, BinaryThresholding, Filling holes). LSTM as a kind of Recurrent Neural Network (RNN) first wasused to tackle the long-term dependencies problem. This problem was due to the fact that ClassicRNNs could not give accurate predictions based on long-term information. Later on, investigationsshowed promising results in mixing the LSTM model within CNN architecture. We usedCNN-LSTM as the model and LSTM was implemented before fully connected hidden layers.Then Adam Optimizer and Cross-Entropy loss function were employed in training the model. Inthis project, we used the dataset provided by M Amir Eshraghi and Mustafa Ghaderzadeh (BloodCells Cancer (ALL) dataset), which includes ۳۲۴۲ Peripheral Blood Smear (PBS) images of ۸۹patients: ۵۱۲ benign, ۹۵۵ Pre-B, ۷۹۶ Pro-B, and ۹۷۹ early Pre-B labeled images (using flow cytometry).Results: Our model reached accuracy, precision, recall, and specificity equal to ۹۶%, ۹۵.۲۱%,۹۵.۱۵%, ۹۸.۶۴% respectively after ۲۲ epochs.Conclusion: This method achieved fairly good results, especially because it can differ ALL subtypeswith such accuracy that in traditional clinical settings can only be done by flow cytometry.Using these methods can reduce human errors, laboratory costs and the time between diagnosisand the start of treatment. In general, the development of such models can assist clinicians in theprocess of initial screening of hematological malignancies, using only a peripheral blood smear.For future work this method can be further expanded on other datasets and for other types of hematologicalcancers such as AML, CML or CLL.

کلیدواژه ها:

Acute Lymphocytic Leukemia ، Deep Learning ، Peripheral Blood Smear ، Classification ، Long Short Term Memory