An Ensemble-Of-Deep Learning Model with Optimally Selected Features for Osteoporosis Detection from Bone X-Ray Images

Abstract

Osteoporosis is a condition wherein bone tissue deteriorates and bone strength deteriorates. Over time, the disease that can lead bone to become more permeable and fragile makes it more prone to breaking. Early recognition of osteoporosis via an X-ray approach will enhance diagnosis and treatment practices, but it will also assist in preventing national economic loss via mass screening and awareness. A novel early osteoporosis diagnosis model is developed for X-ray images in this research work. The following four primary steps are used to construct a unique Osteoporosis detection model in this study: "(a) pre-processing, (b) feature extraction, (c) optimal feature selection, and (d) Osteoporosis detection". Gabor filtering (noise reduction) and histogram equalization are used to pre-process the obtained raw data (X-ray) (quality enhancement). Features such as "Active shape model (ASM), active appearance model (AAM), gray level co-occurrence matrix (GLCM), mean local gradient pattern (M-LGP), and improved median ternary pattern (I-MTP)" are recovered from the pre-processed data. Following that, a new hybrid optimization model chooses the best features from the retrieved features. The cat guided hummingbird foraging algorithm (CGHFA) is a conceptual combination of the basic artificial hummingbird algorithm (AHM) and the cat hunting optimization algorithm (CHOA). The deep learning classifiers in the Osteoporosis detection phase are trained using these ideally selected characteristics. The newly created ensemble-of-classifiers model is used to represent the osteoporosis diagnostic phase. "Quantum deep neural network (QDNN), improved deep convolution neural network (I-DCNN) and recurrent neural network (RNN)" are some of the deep learning classifiers that is employed here. All of these classifiers are trained using the optimal features available. The loss function of DCNN is improved via harmonic mean based cross-entropy function. The final detection performance will be calculated by combining the results obtained from all of these characteristics (by taking the mean). Finally, the effectiveness of the anticipated model is validated by a comparative examination. Accordingly, the detection accuracy attained by the proposed deep ensemble model +CGHFA at Learn_rate=60 is 90.7%, at Learn_rate=70 is 92.14%, at Learn_rate=80 is 93.482% and at Learn_rate=90 is 94.8%, which is higher than the existing models.

Classification of Knee X-Ray Images by Severity of Osteoarthritis Using Skip Connection Based ResNet101

“…The tibiofemoral bones generate erosion in the knee movement, which leads to knee OA [5]. The hyaline cartilage is comprised of cartilage cells that are useful for bone by load allocation and it works for lifelong [6].…”