A MobileNetV۲-based Android Application for Acute lymphoblastic Leukemia diagnoses and classification of its subtypes

Background and aims: B-cell acute lymphoblastic leukemia (ALL) is a common type of cancer,and a conclusive diagnosis involves intrusive, expensive, and potentially harmful diagnostic procedures.However, in many locations, access to accurate ALL diagnostic tools is restricted. Bloodmicroscopic analysis has long been the main ALL screening and diagnosis method. However,there are inherent limits to the microscopic analysis of blood performed by laboratory staff andhematologists. In contrast, image analysis of blood microscopy data using artificial intelligence(AI) approaches has demonstrated encouraging outcomes. This study’s goal was to create and usea well-optimized deep convolutional neural network (CNN) to identify the subtype of ALL byfirst detecting ALL instances in hematogenous.Method: A mobile application was developed with the objective of accurately differentiatingbetween instances of ALL and hematogone cases, utilizing a well-designed and optimized model.The segmentation of images was carried out during the modeling stage using a specialized techniquebased on unique segmentation. The K-means clustering algorithm was applied to segmentthe images and remove irrelevant components, followed by the addition of a mask to the clusteredimages. After evaluating the performance of six popular lightweight CNN architectures (MobileNetV۲,ResNet۵۰, DenseNet۱۲۱, MobileViT, ShuffleNetV۲, and NASNetMobile), the mosteffective model was chosen. The proposed model was then configured and fine-tuned based onthis selected architecture.Results: The proposed model exhibited a remarkable accuracy rate of ۹۹.۹۸%. Based on thisstate-of-the-art model, a mobile application was developed. The mobile application, built uponthe suggested model, effectively differentiated ALL cases from other classes in a real laboratorysetting, demonstrating a ۹۹.۸۸% sensitivity and specificity. This successful implementation establishesthe mobile application as a reliable screening tool for ALL.Conclusion: The mobile application, which incorporates preprocessing and deep learning (DL)algorithms, can be employed by hematologists and clinical specialists as a powerful screeningtool to potentially reduce unnecessary bone marrow biopsy cases and expedite the diagnosis ofALL. This application has the potential to significantly shorten the time required for ALL detectionand diagnosis, thereby improving clinical efficiency and patient care.