Şengül DOĞAN, Burak TAŞCI, Türker TUNCER

Detection of acute lymphocytic leukemia (ALL) with a pre-trained deep learning model

Introduction

Detection of acute lymphocytic leukemia (all) with a pre-trained deep learning model . Detect Acute Lymphocytic Leukemia (ALL) using a pre-trained DarkNet19 deep learning model on peripheral smear images. Achieved 99.94% accuracy, significantly aiding pathologists in diagnosis.

Abstract

Acute Lymphocytic Leukemia (ALL) is a type of cancer caused by immature lymphocytes in the bone marrow. Acute Leukemia is common in both children and adults. It can also cause death if left untreated. Hematologists diagnose ALL by examining the blood and bone marrow. This method used is slow and takes more time. In this study, the diagnosis and classification of the disease was carried out using peripheral smear images with the proposed method. In the proposed method, 99.80% accuracy was obtained by using the DarkNet19 pre-trained model. Then, 1000 features were obtained from Darknet19. 521 of the obtained features were selected with Mrmr feature selection algorithm. The selected features are classified with support vector machines. An accuracy of 99.94% was achieved with the proposed method. The results show that the proposed method can be used as a tool that will certainly assist pathologists in diagnosing ALL and its subtypes.

Review

This paper presents an intriguing approach for the automated detection and classification of Acute Lymphocytic Leukemia (ALL) using peripheral smear images, addressing the time-consuming nature of traditional manual diagnosis. The proposed methodology leverages a pre-trained deep learning model, DarkNet19, for initial feature extraction, followed by a sophisticated feature selection step using the Mrmr algorithm, and finally classification with Support Vector Machines (SVM). The abstract reports highly impressive accuracy rates, initially 99.80% with DarkNet19 alone, which further improves to 99.94% after feature selection and SVM classification. The authors suggest that this method could serve as a valuable tool to assist pathologists in accelerating ALL diagnosis and its subtypes. A significant strength of this work lies in its potential to introduce efficiency and standardization into the diagnostic process for ALL, a critical step given the disease's severity and the need for timely intervention. The hybrid approach, combining the power of deep learning for robust feature learning with a more traditional machine learning classifier after targeted feature selection, is a well-regarded strategy in medical image analysis, particularly when working with potentially complex or limited datasets. The use of a pre-trained model like DarkNet19 is a practical choice, often leading to better performance and faster convergence than training from scratch, especially if the training dataset for ALL images is not exceedingly large. While the reported accuracy of 99.94% is exceptionally high and certainly promising, it also prompts a need for more detailed scrutiny regarding the experimental setup and dataset characteristics. The abstract lacks crucial information about the size, diversity, and source of the peripheral smear image dataset, as well as the specific validation strategy employed (e.g., cross-validation folds, independent test set, patient-level vs. image-level splits). Achieving such high accuracy often warrants careful consideration of potential data leakage or the complexity of the classification task within the specific dataset. Future work should elaborate on these aspects, provide a breakdown of other critical performance metrics like sensitivity, specificity, precision, and recall, and ideally demonstrate the model's robustness and generalizability across varied clinical settings and diverse patient populations, including explicit details on how different ALL subtypes were handled. A thorough comparison with existing state-of-the-art methods for ALL detection would also significantly strengthen the paper's contribution.

Full Text

Comments

You need to be logged in to post a comment.