Leveraging Explainable AI and Optimized Deep Learning for Enhanced Leukemia Detection Using Otsu Threshold Segmentation
R Thanga Kumar, Riddhi Mirajkar, P. Gajendran, Noopor Pandey, Chippalthurthy Gayathri, Mohan Garg · 2024
The identification of aberrant blood cell features is an essential part of the leukaemia diagnosis process, and at the moment, this is accomplished by manual inspection by pathologists who have received training. The morphological characteristics of white blood cells are examined in microscopic pictures as part of this process. The incorporation and implementation of a computerised system vision-based system presents a number of important problems, despite the fact that there are certain machines and chemical-based tests accessible. This stage is responsible for isolating the areas of interest in order to provide an accurate diagnosis. The current work investigates an efficient and composite segmented strategy to improve the effectiveness and efficacy of leukaemia diagnosis. This is done in order to solve the issue that has been raised. For the purpose of this inquiry, the dataset that has been employed is the acute lymphoblastic leukaemia-related image database, often known as ALL-IDB. The Otsu and Kapur approaches, which make use of multiple thresholding, take the pictures through a phase of pre-processing before they are segmented. The use of the concept of enthusiasm-based teaching-learning-based optimisation (LebTLBO) method results in an additional enhancement of the results of segmentation. The suggested LebTLBO approach outperforms the others by a wide margin, reaching an accuracy of $\mathbf{9 8 \%}$. Kapur, MTOtsu, Otsu, and MTKapur are all less accurate than MTKapur. This demonstrates how the LebTLBO approach outperforms the others in terms of error minimisation and result reliability.