Multi‐Class Blood Cancer Malignancy Prediction From Microscopic Blood Cell Images Using ShuffleNetV2

ABSTRACT Accurate identification of blood cancer subtypes from microscopic blood cell images is essential for timely diagnosis and effective treatment planning. Manual examination of blood smears is time‐consuming and subject to observer variability, which highlights the need for automated and reliable diagnostic support systems. While deep learning has shown strong potential in leukaemia classification, many existing approaches rely on computationally intensive architectures that limit practical deployment. This study proposes a lightweight deep learning framework based on ShuffleNetV2 for multi‐class blood cancer malignancy prediction. The model is designed to balance classification performance and computational efficiency in a realistic multi‐class setting. Microscopic blood cell images from five categories; Acute Lymphoblastic Leukaemia, Acute Myeloid Leukaemia, Chronic Lymphocytic Leukaemia, Chronic Myeloid Leukaemia, and healthy samples. A publicly available dataset of 15,000 microscopic blood cell images, evenly distributed across the five classes, was partitioned using a stratified 70:15:15 train–validation–test split to ensure balanced and unbiased evaluation. The images are processed using a standardised preprocessing pipeline and classified through a ShuffleNetV2‐based architecture. Model performance is evaluated using accuracy, area under the curve (AUC), precision, recall, F1‐score, and confusion matrix analysis. Experimental results show that the proposed model achieves an accuracy of 91.91%, an AUC of 0.9938, and a macro‐averaged F1‐score of 0.9176, with a Matthews Correlation Coefficient of 0.9022 and a Cohen's Kappa of 0.8989 confirming strong, chance‐corrected agreement. The model maintains a compact size of approximately 15.26 MB with about 1.27 million parameters and a measured inference cost of only 0.29 GFLOPs. Confusion matrix analysis indicates that misclassifications occur mainly between morphologically similar leukaemia subtypes, most notably Chronic Lymphocytic Leukaemia being confused with Acute Myeloid and Acute Lymphoblastic Leukaemia, while healthy samples are correctly classified in all cases. These results show that ShuffleNetV2 provides an effective and efficient solution for multi‐class blood cancer classification, offering strong diagnostic performance with low computational cost and suitability for deployment in resource‐constrained clinical environments. The results demonstrate that a channel‐efficient architecture can match the diagnostic reliability of substantially larger models while remaining suitable for point‐of‐care and resource‐constrained clinical deployment.