Cancer Cell Classification From Peripheral Blood Smear Data Using the YOLOv8 Architecture
Joao C. S. Nunes, José E. B. de S. Linhares, Miguel Postigo, Daniel Guzmán del Río, Angilberto Muniz Ferreira Sobrinho, Israel Gondres Torné
Abstract
The accurate classification of cancer cells in peripheral blood is essential for the diagnosis of leukemia and has traditionally been carried out by analyzing laboratory images. In this context, the use of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">deep learning</i> techniques facilitates decision-making and speeds up the early diagnosis of the disease, allowing preventive measures to be adopted for the patient. This study explores the application of the YOLOv8 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">deep learning</i> architecture for the classification of cancer cells in blood smear images, due to its ability to perform this task quickly and accurately. The models were trained on two datasets, ALL and C-NMC Leukemia, and evaluated using the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">top-1 accuracy, top-5 accuracy</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">loss</i> metrics. The proposed approach achieved a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">top-1 accuracy</i> of 99.982% and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">top-5 accuracy</i> of 100% on the validation and test subsets of the ALL dataset, with a final <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">loss</i> of 0.02952. For the C-NMC Leukemia <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">dataset</i>, the model obtained 66.897% and 89.612% <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">top-1 accuracy</i> in the validation and test subsets, respectively, while maintaining 100% <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">top-5 accuracy</i> in both, with a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">loss</i> of 0.18289. These results demonstrate the effectiveness of YOLOv8 in classifying cancer cells, especially in the ALL set. However, future strategies such as expanding the data set and fine-tuning the hyperparameters could contribute to better generalizing the model to different data distributions.