Page 71 - AIH-2-3

Artificial Intelligence in Health CNN model for leukemia diagnosis

research contributes to the advancement of automated feature extraction with the CNN architecture, the model
medical image analysis, ultimately aiming to improve becomes more adept at recognizing variations across
patient care and outcomes in leukemia treatment. This different resolutions, enabling it to handle complex and
paper explores the role of multilevel image classification imbalanced medical datasets effectively.
using DL, specifically focusing on the C-NMC dataset. Furthermore, the use of the Tversky loss function in
1.3. Organization this model addresses one of the core challenges in medical
image classification – class imbalance. Medical datasets,
The paper is organized as follows: Section 2 provides a particularly those used for leukemia diagnosis, often suffer
comprehensive review of related work, highlighting key from an uneven distribution of samples across different
advancements in DL for medical image analysis. Section categories. The Tversky loss function, which adjusts the
3 describes the proposed methodology, detailing the CNN trade-off between false positives (FP) and false negatives
architecture, the use of the Tversky loss function, and (FN), ensures that the model remains sensitive to minority
the training process. Section 4 presents the experimental classes, improving overall performance on imbalanced
results, including a performance comparison with other datasets. This is particularly important in the clinical
state-of-the-art models. Finally, section 5 concludes the context, where minimizing FN is crucial for early detection
paper, discussing the implications of our findings and and treatment planning.
potential directions for future research.
Overall, the combination of CNN architecture, multi-
2. Related work scale feature extraction, and the Tversky loss function
presents a robust solution for multilevel classification in
2.1. Review of DL in medical image analysis leukemia diagnosis. By incorporating techniques proven
The application of DL in medical image analysis has successful in other areas of computer vision, this model sets
garnered significant attention in recent years, with a new benchmark for automated medical image analysis,
numerous studies demonstrating its potential to transform offering enhanced diagnostic accuracy and reliability.
diagnostics. This section reviews existing literature on 2.2. Existing solutions for leukemia diagnosis and
the use of DL for leukemia detection and classification, gaps addressed by this study
highlighting key methodologies, findings, and gaps that
this study aims to address. A key strength of this approach Table 1 presents a detailed literature review table
lies in the model’s incorporation of multi-scale features, summarizing the related work involving DL, the C-NMC
a concept widely recognized for enhancing performance dataset, and leukemia, including author details for 15
in other areas of computer vision. Multi-scale features studies from 2020 to 2024.
have shown considerable utility in tasks such as image Figure 2 presents a structured object-oriented model
quality assessment, visual saliency detection, and person for leukemia classification, breaking it down into acute
re-identification. For example, Varga illustrated the and chronic forms, with subtypes such as myeloid and
1
effectiveness of multi-scale orderless pooling of deep lymphocytic variants. One of the key advantages of this
features for no-reference image quality assessment, structure is that it facilitates systematic data representation,
emphasizing the role of feature pooling at different scales making it easier for machine learning (ML) or DL models
to capture both global and local image characteristics. to process and identify patterns across the subtypes. This
Similarly, Li and Yu demonstrated the value of multiscale model provides a clear hierarchy that can be used to label
2
deep features in visual saliency detection, where analyzing and categorize medical data, improving the efficiency and
features at multiple scales helped detect visually significant accuracy of automated disease detection systems.
regions across images. In person re-identification, Chen
et al. utilized multi-scale DL architectures to improve In addition, by organizing the leukemia subtypes into
3
recognition performance by mapping features across a hierarchical object-oriented structure, the relationships
different views and scales, highlighting the versatility of between different types are easier to understand and
manage. This kind of classification allows for a more
this approach.
detailed and accurate analysis of blood samples, which
In the proposed CNN model for leukemia diagnosis, can aid in the differentiation of leukemia types during
multi-scale features are leveraged to detect subtle the diagnostic process. It is particularly useful when
morphological differences between normal and leukemic dealing with large datasets, where the defined subtype
cells. This is critical for accurate subclassification of structure ensures that various forms of leukemia are
leukemia types, as cell morphology can vary significantly appropriately identified, facilitating early diagnosis and
between different subtypes. By integrating multi-scale targeted treatment plans.

Volume 2 Issue 3 (2025) 65 doi: 10.36922/aih.4710