Development of Efficient Schemes for Detecting Hematological Disorders

Abstract

Hematological disorders like Acute Lymphoblastic Leukemia (ALL), Acute Myeloid Leukemia (AML), and Sickle Cell Anemia (SCA) are life-threatening diseases, which severely affect blood cells and the overall health condition as well. Acute Leukemia (ALL or AML) is a fast-spreading blood cancer that starts in the bone marrow and severely damages white blood cells (WBCs). ALL critically affects lymphocytes and produces lymphoblasts, whereas AML seriously damages myeloid cells. Both ALL and AML severely deteriorate the immune system by reducing the ability to fight against infections. On the other hand, SCA critically damages red blood cells (RBCs) and produces sickle cells by replacing healthy hemoglobins with sickle-hemoglobins. It is so dangerous that it reduces the life span of RBCs from 120 days to 10-20 days. Hence, early disease detection and proper treatment are the utmost requirements to save valuable lives and ultimately optimize mortality rates. The microscopic blood-cell analysis is a cost-effective, efficient, less painful, and reliable approach for early hematological disorder detection.
Recent advancements in hematological disease detection systems have aimed to improve the disease diagnosis performance and make the system more accurate and faster as well. Current developments in computer vision, deep and machine learning in medical image analysis make the Computer-Aided Diagnosis (CAD) system more popular for automatic disease detection. However, noise, intensity inhomogeneity, weak edges, overlapping, and touched cells make the segmentation and classification more challenging.
In this thesis, the objective is to develop a more efficient CAD-based system for automatic diagnosis of ALL, AML, and SCA yielding more accurate and more precise results. For this purpose, several models have been developed to investigate the performance. In machine learning-based classification models, we emphasize the improvement of classification ability and are equally concerned about more accurate segmentation.
In Chapter 3, a novel hybrid ellipse-fitting-based segmentation model (HEFSM) is proposed by hybridizing a new geometric ellipse-fitting approach (in which minor and major axes are more accurately evaluated depending upon the residue and suggested residue-offset) with algebraic ellipse-fitting. Hence, it yields superior performance than other ellipse-fitting approaches since it retains the benefits of both approaches. Furthermore, least-square-based HEFSM is computationally less expensive as it is a hybrid of noniterative geometric and algebraic ellipse-fitting approaches. Moreover, bounded opening followed by fast radial symmetry (BO-FRS)-based seed-point detection is recommended to attain more accurate detection.
Though the proposed HEFSM outperforms other ellipse-fitting approaches, its segmentation performance is comparatively less accurate than level set methods (LSMs) as the structures of blood cells are not perfectly elliptical. It may also suffer from low break-down point issues due to least-square-based ellipse-fitting. It inspires us to develop a novel level set evolution-based blood cell segmentation model. A new adaptive weight optimized level set evolution (A WOLSE) is proposed, where the significant contribution is to update the weights of the area- and edge- terms adaptively by optimizing the energy functions. Thus, it results in more precise boundary detection. Nevertheless, a hybridization of AWOLSE with the watershed algorithm is performed to achieve proper segmentation of touched and overlapping cells. Computationally efficient adaptive weight optimization and excellent segmentation performance with more accurate boundary detection make it a preferred segmentation scheme for practical implementation.
In chapter 4, several classification models have been developed for more accurate acute leukemia (ALL or AML) detection using machine learning and deep learning.
First, a hybrid machine learning-based acute leukemia (ALL or AML) detection system is developed by hybridizing Support Vector Machine (SVM) with Random Forest (RF). The efficient AWOLSE-based segmentation model (the best performing segmentation model, as discussed in Chapter 2) is applied to yield more accurate contour detection. In addition, a hybridization of A WOLSE and Watershed algorithm is performed to achieve proper segmentation of overlapping cells. Then, Gray Level Co-occurrence Matrices (GLCM)-based feature extraction followed by Principal Component Analysis (PCA)-based feature selection are employed to select more significant features.
Recent advancements in deep learning motivate us to develop deep learning-based classification systems for ALL or AML detection, which does not need additional segmentation steps, unlike machine learning-based systems. However, classical deep learning models are essential for training the models properly to deliver outstanding performance. On the other hand, the major challenging issue in medical image processing research is the unavailability of huge standard databases. Nowadays, transfer learning is becoming an emerging trend of research in the medical image processing field because of its promising performance in small databases. It inspires us to develop efficient transfer learning-based leukemia detection systems.
Second, a lightweight ShufileNet-based blood cancer detection system is developed whose salient features: pointwise-group convolution and channel shuffling are responsible for making the system faster and efficient.
Third, a new transfer learning-based acute leukemia detection model is developed. It consists of lightweight MobileNetV2-based feature extraction followed by an SVM-based classification. Inverted residual bottleneck structure, depthwise separable convolution, and tunable hyperparameter make the feature extraction effective and computationally efficient as well. In addition, in SVM, the optimization of hyperplane locations boosts the classification performance.
Fourth, a novel hybrid transfer learning framework is developed to more accurately detect acute leukemia by hybridizing lightweight MobileNetV2 with ResNet18. An efficient weight factor is recommended to achieve efficient hybridization. Thus, the integrated advantages of both MobileNetV2 and ResNet18 (inverted residual bottleneck structure, depthwise separable convolution, tunable multiplier, identity mapping, and residual learning) improve the excellence of the system.
Fifth, a novel Orthogonal Softmax Layer (OSL)-based acute leukemia detection system is developed to improve the performance further. The ResNet18-based feature extraction and OSL-based classification make it more effective. Hence, the merits of ResNet (identity mapping and residual learning) are integrated with the merits of OSL ( enhancement of computational efficiency and feature discrimination ability) to deliver promising classification performance.
In Chapter 5, two new Atrous Convolution-based Hybrid DeepLabV3+ Architectures (ACHDAs): ACHDA-I and ACHDA-11, are designed and developed to yield efficient semantic segmentation for more accurate Sickle Cell Anemia (SCA) detection. These two deep learning-based semantic segmentation schemes can yield more efficient anomaly localization in addition to more accurate disease detection. Here, a Modified DeepLab V3+ Architecture (MDA) is recommended in which lightweight MobileNetV2 or ResNet50 is employed as a base classifier rather than a computationally expensive Xception. Moreover, ACHDA-I is developed by hybridizing MDA-1 (MDA with MobileNetV2-based classification and Adaptive moment (Adam)-based optimization) with MDA-2 (MDA with ResNet50-based classification and Stochastic Gradient descent method (SGDM)-based optimization). Hence, it combines the merits of both the classifiers and both optimizers that result in performance improvement. Furthermore, ACHDA-11 is suggested in which only the input image's saturation information (S-component ofHSV color model) is utilized to minimize false-positive and improve performance further. More importantly, these two schemes: ACHDA-I and ACHDA-11, also retain the benefits due to the employment of Atrous convolution and Atrous Spatial Pyramid-based Pooling (ASPP) for feature extraction and implementation of the efficient decoder module for upgrading the segmentation performance at object boundaries.
The experimental results show that the proposed models yield superior performance than the other models, available in the literature, in respective categories. Among the proposed segmentation models, the A WOLSE-based segmentation model outperforms others, whereas the (OSL)-based acute leukemia detection system exhibits the best classification performance in detecting ALL or AML. The proposed semantic segmentation scheme: ACHDA-11, demonstrates the best SCA detection performance.