Body Part Classification from Gabor Enhanced X-Ray Images using Deep Convolutional Models

Large scale availability of archived medical data lends itself to train the Artificial Intelligence (AI) models, which could result in computer aided clinical decision support systems, for enhanced patient care, training of physicians and biomedical research, thereby leading to clinical automation. This work focuses on classification of 22 classes of x-ray images of the UNIFESP dataset. Initial experiments, using features suitable for scale invariance (using Scale Invariant Feature Transform), rotational invariance (using Speeded Up Robust Features) and performance improvement (using Oriented FAST and Rotated BRIEF) were carried out in combination with seven traditional machine learning algorithms, with and without traditional image enhancement techniques. The results were suboptimal due to the class imbalance in the dataset. To address this issue, augmentations using both geometric transformations and generative models were investigated. The augmented images were enhanced using Gabor filter, for brightness and contrast normalization. Thereafter deep learning classification was attempted using five convolutional neural networks namely VGG16, InceptionV3, Resnet50, DenseNet121 , and EfficientNetB3 . Highest accuracy of 86.94% was observed using EfficientNetB3 .