الفهرس | Only 14 pages are availabe for public view |
Abstract Human health can be adversely affected by tumors, inflammations, fractures, and other bone diseases. Medical imaging using skeletal scintigraphy images is one of the most significant and crucial methods to detect bone diseases. Early diagnosis of bone diseases is very important in the treatment phase of the patient. The diagnosis of skeletal scintigraphy images depends on the expert, which requires a lot of experience, time, and effort. Dealing with Egyptian diseases is a major issue in this thesis. Bone metastasis is common in Egypt, especially in Menoufia University Hospital. The diagnosis of bone diseases varies for children, women, and men. Furthermore, the main contribution is to develop a diagnosis algorithm for bone diseases. So, three algorithms are proposed. For enhancing the dark parts of the skeletal scintigraphy images, an adaptive algorithm is proposed. It is built using the Salp Swarm Algorithm (SSA) and a Neutrosophic Sets (NS) with several criteria. The optimum improvement for each individual image is first determined using the SSA algorithm, and the NS algorithm is used to determine the similarity score for each image using the adaptive weight coefficients. For segmenting the dark parts from the bone in the skeletal scintigraphy images, a multi-threshold algorithm is proposed. It segments the skull, the trunk, and the lower limbs separately rather than the whole skeletal scintigraphy image. A sharpness index for each of the three parts is evaluated, and the best threshold is computed using the SSA with a fitness function based on maximizing the Tsallis entropy function for each part. For diagnosing bone diseases in skeletal scintigraphy images, a multiclass with mixed data types algorithm is proposed. It depends on segmented organ images, patient data, and statistical features. Four phases are involved, including pre-processing, feature extraction and selection using the SSA with the age, gender, and organ name of the patient, modeling phase, and classification phase. In the modeling phase, the segmented organs image model relies on three self-attention layers with stacks of three convolution layers, and the selected features of each organ model are composed of Inception-V3 with stacks of three convolution layers and nine inception modules. A single feature map is used to generate the different output layers. The concatenation model is used to combine them and enable the algorithm to learn new features with a dense layer. The classification phase of fully connected layers with the SoftMax activation function is applied to classify each organ into one of three classes: normal, tumor, or inflammation. These proposed algorithms are tested using different measures according to each criterion. The efficiency of the proposed algorithms achieved superior performance. The sharpness index of the enhancement algorithm is 58.84058. The accuracy value of the segmentation algorithm is 96%. The accuracy value of the classification algorithm is 97.5% with a loss value of 0.09. |