Classification of main types of leukemia disease with artificial intelligence-based algorithms on the DNA sequences using digital mapping techniques

A cell is a unit containing structural and functional properties related to life. It has the ability of controlled divisibility. Thus, injured tissues are repaired or dead cells are renewed. However, genetic or environmental parameters can negatively affect this division process. In such a case, malignancies occurrences can happen. Malignancies are described as malignant tumors. It is formed by the uncontrolled proliferation and spread of cells in a particular tissue or organ. Cancer is a malignancy that occurs with the uncontrolled proliferation of cells in any part of the body. With this disease, the body's regular working mechanism gradually deteriorates and becomes irremediable if an early diagnosis is not provided. This process is important in terms of determining the type of disease and deciding the treatment methods. In this study, leukemia malignancy, which is among the cancer types, was examined. Because leukemia is the most common type of cancer. It can be seen in all age groups. According to the course of the disease in the person, it is examined in two groups as acute or chronic. Early diagnosis is essential for a long life expectancy, especially in patients with acute leukemia, where the disease appears suddenly and spreads rapidly throughout the body. However, about 40 years ago, a type of leukemia, defined by doctors as mixed lineage leukemia (MLL), was discovered in certain types of patients diagnosed with ALL and AML, which are the main subtypes of leukemia. After the discovery of the MLL leukemia type as a new species, it is thought that the leukemia disease still contains new sub-formations that can be discovered. Clearly distinguishing the subtypes of leukemia cancer, which is considered in a wide framework, will ensure the implementation of the correct treatment protocol together with the realization of the correct diagnosis process. On the other hand, it is also vital in terms of evaluating the response to treatment and clarifying the protocol to be applied for subsequent treatments. Peripheral blood smear, bone marrow aspiration, bone marrow biopsy, immunophenotyping, and various imaging and tests are used to reach definitive results in the medical world. However, these methods sometimes involve uncertain situations. For example, the manual evaluation process of the peripheral blood smear method depends on the doctor's knowledge, experience, physical intensity and mental fatigue. Cancerous tissue may not be present in the part taken by the biopsy method from a person with cancer. The increase in lymphocyte cell type, which is an indicator in the diagnosis of leukemia, may also increase as a result of hepatitis viruses and brucella disease. The success of applying imaging methods depending on various medical parameters or the evaluation of the achieved imaging by doctors may not produce standard output. Therefore, molecular and cytogenetic evaluation of genetic-based malignancies will clarify the accuracy of the diagnosis. However, recent advances in genome technology have led to an increase in DNA sequences. For this reason, it has become difficult to manually interpret data accurately and quickly in line with cytogenetic analysis. On the other hand, microarray technology, which offers the possibility of molecular evaluation, was evaluated. However, the analysis process of microarray technology, which is a source of hope in cancer research, offers a high dimension because it is provided through features that represent genes. This situation, known as the curse of dimensionality in machine learning, can result in computational burden and noise that produces wrong decisions. In the scope of all these methods, the number of treatment approaches planned to be applied to the person may cause anxiety in the patient within the framework of the cost parameter. At the same time, unnecessary radiation exposure and workload and time loss for doctors are other negative outcomes. In this study, a computer-aided structure was built in 7 separate sections in order to provide an early diagnosis with fewer medical methods. With this structure, which was designed to give an idea to doctors, leukemia disease was analyzed for 7 different purposes. In the first stage, where cytogenetic evaluation is achieved, BCR-ABL genes, which are an important indicator in the diagnosis of ALL and CML malignancies, which are the main types of leukemia, were analyzed. BCR-ABL genes supplied from the NCBI dataset were digitized using mapping techniques. Then, fourier transform and short-time fourier transform methods were used in the scope of the signal processing approach to detect exon regions containing critical information about life on the digitized array. Spectral density information obtained from DNA sequences with signal processing methods was classified with EfficientNetB7 transfer learning architecture. With the visual data expressing the spectral density of different DNA base lengths included in the training dataset, the accuracy rate reached on the test dataset is 100%. The accuracy rate achieved in the test dataset is between 50% and 70% for visual data expressing the spectral density of DNA base length in a certain range in the training dataset. In the second stage, where cytogenetic evaluation is achieved, malignancies were differentiated by using BCR-ABL genes obtained from ALL and CML patients via the NCBI dataset. This process took place in two separate parts. In the first part, the spectral density information of the digitized DNA sequences was reflected to the spectrograms with short-time fourier transform and continuous wavelet transform methods. Then, the spectrograms were classified with the DGCNN method, which provides temporal and spatial information extraction. In this part, where a classification based on nucleotide base length takes place, the maximum accuracy rate achieved for both short-time fourier transform and continuous wavelet transform methods was obtained as 75%. In the other part, a study independent of nucleotide base length was performed. In this direction, statistical and structural features were extracted from digitized DNA sequences using Shannon entropy-based mapping technique. Then, these features are given as input to the adaptive fuzzy logic algorithm. With this method, which is thought to be suitable for the fuzzy configuration structure of DNA, a accuracy rate of 80% was achieved. In the third stage, where molecular evaluation is achieved, the microarray dataset of ALL and AML malignancies, which are the main types of leukemia, obtained by the bioinformatics laboratory, was used. In the first step, potential genes were selected from the high-dimensional microarray dataset using the particle swarm optimization algorithm, ant optimization algorithm and whale optimization algorithm. Then, the spectral density information of the selected genes for each optimization algorithm was reflected in the spectrograms with the continuous wavelet transform method. Finally, spectrograms matching ALL and AML categories were classified by the DGCNN method. The maximum accuracy rate was obtained as 93.33% as a result of the classification of the potential genes selected using the ant optimization algorithm with the DGCNN method. In the fourth stage, where molecular evaluation is achieved, a microarray dataset of ALL, AML and MLL leukemia types supplied by the bioinformatics laboratory was used. Genes selected from the high-dimensional microarray dataset using the whale optimization algorithm were given as input to the LSTM neural network architecture. The average accuracy rate in 100 different datasets created with different seed values of the LSTM architecture, which has the ability to keep dynamic with the actions of remembering the relationship of the selected genes with each other, was found to be 89.883%. In the fifth step, where pathological evaluation is achieved, ALL-IDB1 and ALL-IDB2 peripheral blood smear images obtained from the Acute Lymphoblastic Leukemia Image Database were used. The training of this study, which produces a real-time output for the detection of blast cells and the production of count information, which is an important indicator in the detection of ALL malignancy, which is the main type of leukemia, was carried out by the YOLOv4 algorithm. The accuracy rate of this system, which produces an output within 3-4 seconds for the input given to the model, was found to be 98.87%. In the sixth stage, where molecular and immunophenotypic evaluation is achieved, microarray datasets of T-ALL, B-ALL and T-LL malignancies related to lymphoblast cell types obtained by the bioinformatics laboratory were used. First, potential genes selected from high-dimensional microarray datasets using the whale optimization algorithm were classified with the ANFIS structure. Thus, fuzzy inferences were obtained. Then, the parameter optimization of the ANFIS structure, which combines the inference power of fuzzy logic and the learning power of artificial intelligence on data, is provided by ABC and PSO optimization algorithms, respectively. Finally, 3 different fuzzy outputs from ANFIS, ANFIS+ABC and ANFIS+PSO structures were classified by logistic regression algorithm within the scope of the ensemble learning approach. The achieved accuracy rate is 86.6%. In the seventh stage, where pathological evaluation is achieved, digitized peripheral blood smear images (basophil, eosinophil, lymphocyte, monocyte, and neutrophil white blood cells) were supplied through the Raabin Health dataset. The proposed YOLOv5x approach and hybrid structure were presented to improve the recognition rate of lymphocyte cell type, which is an important indicator in the diagnosis of leukemia disease. With the proposed YOLOv5x approach, an improvement of 0.133 and 0.006 was achieved for the recognition rate for lymphocyte recognition rate and overall detection accuracy, respectively, while an improvement of 3.44% and 14.7% was achieved for the hybrid structure. In this study, a wide analysis including cytogenetic, molecular, molecular and immunophenotypic, pathological evaluations related to the diagnosis and treatment process of leukemia, which has different subtypes, was performed with computer-aided systems. A decision support system was built with 7 separate evaluations presented on leukemia cancer, which is a current research area and is still being studied for the discovery of non-invasive or alternative approaches. It is aimed to realize successful treatments that start with accurate and early diagnosis through this system, which is designed to give an idea to doctors for leukemia, which reduces the quality of life of the person, causes loss of workforce loss and is life-threatening.